A High Throughput Assay for DNA Replication Inhibitors Based upon Multivariate Analysis of Yeast Growth Kinetics

Abstract

Mcm2-7 is the molecular motor of eukaryotic replicative helicase, and the regulation of this complex is a major focus of cellular S-phase regulation. Despite its cellular importance, few small molecule inhibitors of this complex are known. Based upon our genetic analysis of synthetic growth defects between mcm alleles and a range of other alleles, we have developed a high throughput screening (HTS) assay using a well-characterized mcm mutant (containing the mcm2DENQ allele) to identify small molecules that replicate such synthetic growth defects. During assay development, we found that aphidicolin (inhibitor of DNA polymerase alpha) and XL413 (inhibitor of the DNA replication-dependent kinase CDC7) preferentially inhibited growth of the mcm2DENQ strain relative to the wild type parental strain. However, as both strains demonstrated some degree of growth inhibition with these compounds, small and variable assay windows can result. To increase assay sensitivity and reproducibility, we developed a strategy combining analysis of cell growth kinetics with linear discriminant analysis (LDA). We found that LDA greatly improved assay performance and captured a greater range of synthetic growth inhibition phenotypes, yielding a versatile analysis platform conforming to HTS requirements.

Introduction

Automated high-throughput screening (HTS) has revolutionized the discovery of both chemotherapeutics and molecular probes ¹. Early screening efforts often utilized purified systems combined with biochemical assays to identify enzymatic inhibitors. For example, this approach identified agents that block ongoing DNA replication by specifically targeting DNA polymerase (e.g., aphidicolin ²), topoisomerase ³, DNA ligase ⁴, and the DDK kinase CDC7/DBF4⁵. Such DNA replication inhibitors have broad applicability in antiviral and antitumoral chemotherapy ⁶.

As many potential drug targets lack robust and easily assayable enzymatic activities, biochemical approaches have limitations in HTS. For example, few identified inhibitors block the early stages of DNA replication, as many such initiation factors function as multi-subunit complexes that are difficult to purify in sufficient quantity (e.g., ORC and MCM2–7) or lack enzymatic activity in isolation (CDC45 and the GINS complex) ⁷. Such factors are potentially very important drug targets, as their inhibition would be expected to block unwanted DNA replication before it starts, providing an alternative and perhaps less deleterious therapeutic avenue compared to current inhibitors that target on-going DNA replication. In particular, the replicative helicase Mcm2–7 is an unusually promising chemotherapeutic target that is involved in key aspects of both initiation and elongation ^{8, 9}.

The discovery of selective Mcm2–7 inhibitors has been challenging. One initial attempt utilized a low-throughput candidate approach combined with an in vitro biochemical assay to measure DNA unwinding ¹⁰. Several of the inhibitors identified by this assay, including the fluoroquinolone ciprofloxacin, demonstrated low to moderate specificity. However, the assay lacked scalability, as the Mcm2–7 complex is difficult and expensive to purify in sufficient quantities for HTS.

In principle, such limitations can be addressed using a cell-based HTS approach ¹¹. One particularly fruitful method is based upon synthetic growth inhibition ¹². As defined genetically in yeast and bacteria, synthetic growth inhibition occurs when two non-allelic mutations, which individually cause little or no growth defect relative to the corresponding wild type alleles, synergistically cause lethality or sickness when combined in a double mutant strain ¹³. As such genetic interactions in lower eukaryotes are often evolutionarily conserved in human cells ¹⁴, synthetic growth inhibition has been adapted to human cell HTS to identify small molecules that preferentially inhibit a particular cancer cell line but not the corresponding “wild type” cell line¹⁵. Such “hits” likely recapitulate a synthetic growth defect between the target mutation and the gene product being inhibited by the small molecule ^{12, 15}.

Although synthetic growth inhibition was originally studied using null alleles in nonessential genes, similar synergistic growth defects can be observed with certain combinations of non-null mutations in essential genes ¹⁶. Some such non-null alleles that either reduce gene dosage or generate dominant negative alleles likely contribute to human disease. For example, human tumors commonly contain mutations in the genes that encode the Mcm2–7 replicative helicase; some such mutations encode altered gene products while others alter Mcm2–7 dosage ⁸. Although the contribution that such mcm alleles may make to human disease is poorly understood in most cases, a dominant negative mcm allele has been identified that causes cancer in mice (mcm4chaos3 ¹⁷), while experimentally altering Mcm dosage in human tissue culture cells causes genome instability ¹⁸.

To both aid mechanistic studies as well as facilitate potential chemotherapy, novel inhibitors that target the initiation stage of DNA replication are needed. To overcome the biochemical challenges stated above, inhibitors could be identified using a cell-based HTS that utilizes a synthetic growth inhibition assay measuring differential sensitivity of mutant vs. wild type strains to small molecules. However, a general problem of synthetic growth inhibition screening is that putative “hits” will likely also cause detectable growth inhibition of the wild type cell line, thus increasing the difficulty of identifying inhibitors by reducing the statistical growth difference observed between test and control strains.

Developing a better analysis pipeline for synthetic growth inhibition screening would save considerable time and money. Toward this end, we have developed a direct, non-biased differential growth assay based upon optical density changes to identify inhibitors that target a well-characterized Mcm2 mutant (mcm2DENQ). Unlike most other mcm alleles, the mcm2DENQ allele has been extensively characterized both biochemically ¹⁹ and genetically²⁰, and generates the types of genome instability that are commonly observed in human cancers ²¹. To simplify development, this assay utilizes the budding yeast S. cerevisiae, which is both the premier system for studying eukaryotic DNA replication ²² as well as an organism with proven tractability for use in chemogenomic screens (e.g., ²³). Side-by-side comparison of an mcm2DENQ mutant to its wild type isogenic parent is used as the criterion to identify “hits” during primary screening. While conducting a limited candidate screen, we identified two proof-of-concept compounds (aphidicolin and XL413) that meet our criterion and validate our screening premise. Using these candidate compounds, we compared various approaches to analyze and score differential growth of mutant and wild type strains, and developed a sensitive multi-parametric screening algorithm based on linear discriminant analysis (LDA ²⁴) that utilizes the entire growth kinetics of both strains. Using the LDA approach, we find that our assay meets HTS performance criteria and provides a viable analysis pipeline for the quantification of synthetic growth inhibition screening data.

Materials and Methods

Yeast strains and growth conditions.

Strains and plasmids used in this study are described in detail in Supplemental Materials. For assay development, we used a yeast strain in the SEY6210 background (MRY548) that contains deletions of three key multidrug transporter (MDT) genes (hereafter referred to as UPY1325 or Δ3xMDT WT; kindly provided by Scott Moye-Rowley ²⁵); all other strains used were in the W303 background ²⁶. Strain and plasmid manipulations utilized standard methods²⁷. Strains were grown in YPD at 30^oC with agitation unless otherwise noted. To maintain the viability of strains containing deletions of the essential genes MEC1 or RAD53, a mutation of SML1, a known suppressor of this lethality, was added as indicated ²⁸. The sml1 mutation increases cellular deoxynucleotide triphosphate levels, and by itself lacks a noticeable growth phenotype ²⁸. Addition of the mcm2DENQ mutation to strain MRY548 (resulting strain UPY1355 or Δ3xMDT 2DENQ) utilized a standard two-step gene replacement approach²⁷, and the presence of the mcm2DENQ mutation was confirmed by Sanger sequencing (Genewiz) following colony PCR.

Standard microtiter assay.

Wild type and mcm2DENQ test strains were grown as 5 ml overnight cultures in YPD. The next morning, cultures were diluted 1:25 in pre-warmed YPD and growth monitored by spectrophotometric turbidity measurements at 600 nm (Abs₆₀₀). Diluted cultures were re-grown with shaking to log phase (Abs₆₀₀ 0.5 – 0.7) and diluted in YPD to an Abs₆₀₀ of 0.1 – 0.15. 40 μl of these cultures were used to seed the wells of a 384 well clear bottom microtiter plate (Greiner 781091) using an 8 channel multipipette. To enable direct comparisons, both strains were plated on the same 384 well plate. Microtiter plates were sealed with transparent film (Perkin Elmer TopSeal-A, cat# 6005185), incubated at 30^oC or room temperature in an Envision plate reader (Perkin Elmer) without shaking, and Abs₆₀₀ read every hour for 24 hours, with a 10 sec. shake cycle before each read.

For drug treatments, a 96 well plate with serial dilutions of test compounds in YPD containing 1.9% DMSO was prepared, and 15 μl of the resulting solution transferred in duplicate to cells using a Biomek 2000 liquid handler to yield ten-point, two-fold gradients of test agents in quadruplicate and 32 wells of vehicle controls (final [DMSO] = 0.5%). Unless otherwise noted, the Abs₆₀₀ at the time of plating (t=0) was subtracted from each well read, and background corrected values were divided by the Abs₆₀₀ of the respective strain grown in YPD with vehicle only to calculate percent of control = (Abs_600(well) – Abs_{600(well at t=0)}) / mean(Abs _{600(vehicle controls)} - Abs _{600(vehicle controls at t=0)}) * 100. In some cases, data were plotted as percent growth inhibition = 100-(percent of control). Quadruplicate wells of each time point and treatment concentration were averaged, and the results were plotted (mean ± S.D. unless otherwise noted) as a function of incubation time or drug concentration, as appropriate.

Analysis of yeast growth curves by fitting to a modified Gompertz equation.

Growth kinetics were analyzed by a modified Gompertz equation ²⁹ to calculate maximum cell expansion (A), slope (μ), and lag time (λ). Because the Gompertz equation uses a logarithmic scale for cell expansion, well normalization was performed as ln(Abs_600(well) / Abs_{600(well at t=0)}).

Multivariate analysis.

Performance of HTS assays is typically judged by Z’ factors ³⁰ calculated on a single assay parameter. For multiparametric assays much of the information contained in the dataset is unutilized, and single parameters, especially from inherently variable assays or assays with small assay windows, often fail to reach HTS performance (i.e., Z’ factors >0.5). We therefore adapted a statistical method (LDA) that seeks to maximize the discriminant of two classes by projecting multiple parameters to low dimensional parameters ²⁴, to integrate multiple readouts into the Z’-factor as previously described ^31–34. We first calculated LDA values by combining the three Gompertz parameters (A, μ, and λ) from yeast growth curves, using the LDA function included in the MASS package of the R statistical software program (https://cran.r-project.org). Details of the method, together with a sample R script, are provided in Supplemental Material. Z’ factors were then calculated as described ³⁰ using LDA values from WT and 2DENQ strains treated with equal concentrations of test agents (aphidicolin or XL413) as minimum (MIN) and maximum (MAX) controls, respectively. To avoid confusion in the interpretation of figures, it is important to note that 1) LDA reverses the MAX and MIN values (i.e., 2DENQ became positive and WT became negative), and 2) LDA values for the two classes by definition have the same values but with different signs, resulting in a (meaningless) S/B ratio of −1, which we, therefore, did not report.

Results

Identification of synthetic lethal interactions involving the mcm2DENQ allele.

The Mcm2–7 replicative helicase is an ATPase formed from six different and essential subunits historically numbered Mcm2 through 7. The mcm2DENQ mutation is a substitution of two acidic amino acids in the universally conserved Walker B ATPase motif of Mcm2 with their amide counterparts (DE -> NQ, ¹⁹). The corresponding yeast mutant retains good cell growth and viability, but substantially blocks activation of the S-phase replication checkpoint cascade immediately upstream of the effector kinase (Rad53 in yeast), underscoring the functional role of Mcm2–7 in this quality control cascade ²⁰.

During our studies, we noticed that the mcm2DENQ allele causes a synthetic growth defect when combined with viable alleles of various checkpoint genes. As Mcm2 is an essential gene, our analysis used a conditional system so that the test strains can be propagated under permissive conditions, while synthetic-lethal phenotypes can be scored under restrictive conditions. Toward this end, we used a plasmid-swap system (Supplemental Material). In short, the haploid test strain contains a wild type copy of the target Mcm gene expressed from an unstable plasmid additionally containing the selectable marker URA3; this plasmid complements a recessive mutant mcm test allele (usually mcm2DENQ) that is also present in the cell. The conditional nature of the system depends upon the growth media: Growth on non-selective media provides permissive conditions by supporting the presence of the wild type Mcm plasmid; while growth on media containing 5-Fluoro-orotic acid (5-FOA) provides restrictive conditions by selecting for cells that have lost the wild type Mcm plasmid, and thus uncover the mutant mcm phenotype. Additional strains that substitute either a wild type Mcm gene or an empty vector for the mutant test allele serve as positive or negative controls. To test synthetic lethality with the mcm2DENQ allele, additional test mutations are incorporated into the strain as indicated.

Using this system, we found that the mcm2DENQ allele demonstrates synthetic genetic interactions when combined with deletion alleles of the S-phase checkpoint genes MRC1 (= human Claspin), TOF1 (= human Timeless), CSM3 (= human Tipin) or RAD9 (human orthologs include BRCA1 and 53BP1), as well as the checkpoint kinases MEC1 (= human ATR) and RAD53 (=human CHK2) (Figure 1 A and B ³⁵). As shown, the mcm2DENQ mutant, as well as each mutant in these checkpoint genes, individually supports cellular viability to nearly wild type levels in the presence of 5-FOA. However, when the mcm2DENQ allele is combined with these alleles under restrictive conditions, growth is either eliminated (e.g., when combined with deletion alleles of MRC1, TOF1, CSM3, and MEC1) or considerably decreased (e.g., when combined with deletion alleles of RAD9 and RAD53). Thus, the mcm2DENQ allele demonstrates significant synthetic growth defects when combined with a variety of checkpoint alleles.

Figure. 1. Synthetic growth defects involving mcm alleles.

Permissive conditions correspond to growth on YPD, while restrictive conditions that test for synthetic growth defects correspond to growth on 5-FOA media (Supplemental Materials). In the diagram that accompanies each panel, genes written on the circumference of the plate correspond to chromosomal test alleles, while the mcm alleles listed within specific plate sectors correspond to those present on the indicated TRP⁺ test plasmid. A. Indicated strains either lack a chromosomal test allele (wild type (UPY110)) or contain a deletion of the checkpoint mediator genes mrc1Δ (UPY428.1), tof1Δ (UPY631), csm3Δ (UPY632) or rad9Δ (UPY421) as indicated. The TRP1⁺ test plasmids used are Mcm2 (positive control, +2, pUP197) or mcm2DENQ (test allele, +2DQ, pUP199). B. Synthetic lethality among mcm2DENQ, mec1Δ and rad53Δ. The wild type strain (UPY629) is similar to that used in (A) but contains an additional chromosomal deletion of SML1 (Materials and Methods). The indicated strains either lack a chromosomal test allele (wild type (UPY629)) or contain a deletion of the checkpoint kinase genes mec1Δ (UPY1125) or rad53Δ (UPY UPY1124) as indicated. The TRP1⁺ test plasmids used are the same as in A) plus addition of the empty vector control (Vec, pUP169). C. Synthetic lethality between the mcm2–1 and checkpoint mediator alleles. Genetic interaction was tested identical to A), except that TRP1⁺ tester plasmids that encode mcm2–1p (pUP1033) was substituted for the mcm2DENQ plasmid. D. Synthetic lethality between the mcm2DENQ and mcm4RA alleles. The test strain (UPY1597) is similar to UPY110 (A., above) but additionally contains the mcm4RA allele. The following TRP1⁺ test plasmids are used: empty vector (vec, pUP169), Mcm2 (+2, pUP197) or mcm2DENQ (+2DQ, pUP199). For each treatment, the results for two independent colonies are shown.

We next tested the generality of these observations by examining an additional mcm allele, mcm2–1.³⁶. Similar to the mcm2DENQ allele, the hypomorphic mcm2–1 allele causes synthetic sickness when combined with the previous checkpoint alleles under the restrictive condition (Figure 1 C).

Finally, we examined if mcm alleles in different Mcm genes can cause synthetic lethality. Using our plasmid swap system, we examined genetic interactions between the mcm2DENQ allele and the mcm4RA allele (Figure 1 D), a previously characterized viable arginine to alanine substitution allele of the arginine finger ATPase motif of Mcm4 ³⁷. We note that this combination of different mcm alleles also demonstrates synthetic lethality.

In summary, synthetic lethal interactions appear to be a common and general property of Mcm mutants. Below, we leverage the mcm2DENQ mutant to develop a cell-based HTS to identify small molecule inhibitors that mimic these synthetic growth interactions, potentially providing an avenue to identify novel replication checkpoint inhibitors or inhibitors of the Mcm complex.

Assay design and validation.

First, we built appropriate test strains. In addition to their cell wall, yeast contain an unusually large number of multidrug transporters (MDTs), making them naturally resistant to many small molecules ³⁸. We found that a strain deleted for three specific MDT genes (PDR5, YOR1, SNQ2) is sensitive to an inhibitor identified in our prior low-throughput screen (ciprofloxacin), as well as a large variety of other small molecules ²⁵. The mcm2DENQ mutation was added to this triple transporter-defective strain. These two strains differ only in the absence (Δ3xMDT WT, UPY1325) or presence of the mcm2DENQ allele (Δ3xMDT 2DENQ, UPY1355) and unless explicitly stated served as our test strains for the following experiments.

Next, we used Abs₆₀₀ as a measure of cell growth to adapt our genetic synthetic growth inhibition assay into a 384 well format suitable for HTS. To determine the relationship between culture density and optical absorbance, liquid cultures were diluted to the desired density, and parallel Abs₆₀₀ readings from a spectrophotometer and a plate reader were compared. Absorbance measurements performed in cuvettes were linear up to an Abs₆₀₀ = 0.5 (corrected for media background), and correlated with absorbance measurements in 384 well plates (Figure S1 A). This result documented that starting cell densities could be accurately predicted and established by appropriate dilution of yeast stocks.

Because yeast growth and survival in liquid culture are sensitive to changes in environmental conditions ³⁹, we next established the growth characteristics of these strains. Strains were seeded at different starting densities (10 wells per condition) and incubated at 30^oC. Abs₆₀₀ was recorded at the start of the experiment (t=0) and at every hour for 24 hours (t=24). As expected, lag time varied with seeding density. When plated at higher starting densities, the mcm2DENQ strain showed a slightly longer lag phase than the wild type strain; however at low seeding densities (Abs₆₀₀ 0.1), both strains demonstrated nearly identical growth kinetics and reached identical densities at stationary phase (Figure S1 B).

We next ascertained uniform cell expansion by comparing Abs₆₀₀ changes of two full 384 well microplates containing the wild type and mcm2DENQ strains in alternating rows. These plates were kept at 30^oC in the Perkin Elmer Envision multilabel reader, or alternatively in a cell culture incubator and read manually at multiple time points over 24 hours. In both cases, cell expansion was uniform across the plate (Figure S1 C). Based on this data, final assay conditions comprised a low and reproducible seeding density (Abs₆₀₀ = 0.1 – 0.15 by cuvette measurement) and a 24 h incubation time at 30^oC with hourly reads on the Envision, which resulted in uniform cell expansion for both strains and captured a full growth curve over the duration of the experiment.

Identifying model compounds.

To both validate our screening approach and develop an appropriate analysis platform, we required a positive control compound that specifically inhibits the mcm2DENQ strain relative to the wild type strain. Previous results indicated that the mcm2DENQ strain is hypersensitive to the ribonucleotide reductase inhibitor hydroxyurea (HU) ²⁰. Surprisingly, although we found that HU on agar plates caused demonstrable growth defects in the mcm2DENQ strain relative to the corresponding wild type strain, in liquid culture we were unable to statistically recapitulate this defect (Figure 2)

Figure 2. Growth inhibition of selected agents against wild type and mcm2DENQ strains in MDT-replete (W303) and MDT-deleted (Δ3xMDT) strain backgrounds.

For assay trials, compounds were dissolved in DMSO (with the exception of hydroxyurea and XL413, which were dissolved directly in YPD, and cisplatin, which was dissolved in 0.9% saline) and diluted into YPD to their maximal solubility. Ten-point, two-fold concentration gradients were generated by serial dilution into YPD containing 1.9% DMSO, and transferred with a Biomek 2000 liquid handler to assay plates containing an optimized starting density of cells (inhibitor assay concentrations are shown in Table S1). Y-axis shows percent growth inhibition as defined in Materials and Methods using background subtracted Abs₆₀₀ values relative to cells that received vehicle only. All data points were at the highest concentration of compound tested (with the exception of 4-nitroquinoline 1-oxide, which is shown at its IC₉₀ (5 μM) and doxorubicin, which interfered with Abs₆₀₀ readings and was tested at the highest concentration that did not cause assay interference (150 μM). Each data point is the average ± S.D. of 1–7 independent experiments, performed in quadruplicate.

We therefore acquired and tested a panel of small molecules that either cause specific types of DNA damage or inhibit known replication or checkpoint proteins (Table S1). To validate the importance of using MDT delete strains for screening, both the above-described wild type and mcm2DENQ strains lacking the three MDTs (UPY1325 and UPY1355; hereafter referred to as Δ3xMDT background), as well as corresponding wild type and mcm2DENQ strains containing these MDTs (i.e., UPY464 and UPY1062, hereafter referred to as W303 background), were assayed in parallel for direct comparison (Table S2).

This pre-screen yielded some important insights into susceptibility of MDT-deleted yeast to small molecules and their ability to selectively inhibit growth of mcm2DENQ strains. Many compounds lacked apparent toxicity in either strain background, consistent with the known resistance of yeast to small molecules (Figure 2). For compounds that did show activity, there appeared to be a trend to higher potency in the Δ3xMDT background (Figure 2, compare open gray symbols with closed black symbols), supporting our rationale to use MDT deletions to increase hit rates during HTS ²⁵. Two agents were toxic but neither selective for the Δ3xMDT background nor the mcm2DENQ mutation (MMS, 4-nitroquinoline 1-oxide), and another two compounds (camptothecin and etoposide) showed selectivity for the Δ3xMDT background but not the mcm2DENQ mutation.

However, three agents (doxorubicin, aphidicolin, and XL413) satisfied our criterion for a synthetic growth defect “hit”; they selectively inhibited the Δ3xMDT 2DENQ strain relative to the wild type control (Figure 2). Each of these compounds targets specific replication factors that are either needed for Mcm2–7 activation (XL413 targets the DNA dependent kinase Cdc7 ^{5, 7}), that physically interact with Mcm2–7 (aphidicolin inhibits DNA polymerase alpha ^{2, 40}), or that are specifically needed for replication fork progression (doxorubicin inhibits DNA topoisomerase II probably through DNA intercalation ²²), suggesting that these inhibitors uncovered additional, previously unknown, synthetic growth defects involving the mcm2DENQ allele. In addition, the observed differential growth inhibition was not unique to the particular lot of chemicals, as identical results were obtained using doxorubicin, aphidicolin, and XL413 from alternate sources.

Optimizing assay conditions by using the kinetics of aphidicolin and XL413 growth inhibition.

Because at the two highest concentrations doxorubicin’s red color interfered with turbidity measurements at Abs₆₀₀ (making the compound appear less potent; data not shown), we used aphidicolin and XL413 to establish conditions that maximize the difference between the MDT-deficient wild type and mcm2DENQ strains.

To optimize assay windows, we first examined the kinetics of cell growth inhibition (Figure 3 A). Qualitatively, cell growth differences between the two drugs were readily apparent. Aphidicolin appeared to elicit a more sustained growth inhibitory effect whereas XL413 was more transient. Furthermore, the synthetic growth defects by XL413 occurred over a narrower time window as compared to aphidicolin (Figure 3 B). Dose-response determinations indicated that drug effects were most pronounced during active growth (Figure 3 B); at both early (6 hours, early logarithmic growth) and late time points (24 hours, stationary phase), overall toxicity as well as differences between the two strains were low. For these compounds, we found an apparent optimum assay window around 14 hours after the start of the experiment, with an experimentally usable window extending from 12 to 18 hours.

Figure 3. Synthetic growth inhibition by small molecules is concentration- and time-dependent.

UPY1325 (Δ3xMDT WT) and UPY1355 (Δ3xMDT 2DENQ) cells were treated in 384 well plates with ten-point, two-fold concentration gradients of aphidicolin or XL413 and absorbance was measured every hour for 24 hours. A. Growth inhibition kinetics document time- and concentration-dependent differences in WT and mcm2DENQ strains. Y axis shows raw Abs₆₀₀ values not corrected for background. B. Dose-responses at selected time points indicate the optimal separation of WT and mcm2DENQ is growth phase dependent (6 h, early log phase; 14 h, late log phase; and 24 h, stationary phase). Y axis shows percent of control growth calculated as described in Materials and Methods. Data in (A) are the averages ± S.D. of quadruplicate wells from a single experiment that has been repeated four times. Data in (B) are the averages ± S.D. of four independent experiments performed in quadruplicate. Where error bars are missing they are smaller than the respective symbols.

We next examined the effect of incubation temperature on growth inhibition. Although it was convenient to incubate single plates in the plate reader during assay development, this strategy poorly accommodates the handling of multiple plates needed during actual library screening. Thus, we repeated our assay in the presence of various concentrations of aphidicolin and incubated the plate at room temperature (i.e., ~24^oC). Although room temperature incubation slightly retarded the rate of strain growth, comparison of these results to those collected at 30^oC revealed little or no difference on either final cell densities or compound potency at the 14 hour time point (Figure S2). Thus, as modest temperature differences did not impair our ability to detect aphidicolin’s inhibition kinetics and potency, room temperature incubation was incorporated into our standard assay procedure.

Developing a robust analysis platform to achieve HTS assay performance.

As our results indicate, growth inhibition with aphidicolin and XL413 is not an “all-or-nothing” situation that occurs at a specific time point, but rather is best observed when the entire growth kinetics of the assay is considered. Thus, although the 14 hour time point reproducibly demonstrated the maximal growth difference between the wild type and mcm2DENQ strains under our test conditions, it is unclear if other “hits” observed during actual screening will demonstrate either identical kinetics or comparable levels of growth inhibition. To increase assay sensitivity, reproducibility, and ability to accommodate phenotypic variation among different potential inhibitors, we developed an analysis platform that would capture the complete growth kinetics of the experiment.

To develop this multi-parametric approach, we note that the growth kinetics of single-celled organisms can be described by three parameters (illustrated in Figure 4 A): maximal cell expansion (A, defined as the asymptote to the maximal Abs₆₀₀ reached at stationary phase), maximum rate of expansion (μ; defined as the tangent to the inflection point of the growth curve), and lag time (λ; the time it takes for the culture to enter logarithmic growth, defined by the x-axis intersect of the tangent to the inflection point). To calculate these three parameters, we fitted our growth data to a modified Gompertz equation ²⁹ (Figure 4 A). To acquire enough data points to derive meaningful performance statistics (i.e., Z’-factors ³⁰), we seeded half of the wells on a 384 well plate with 3x MTD WT and the other half with Δ3xMDT 2DENQ. All wells were treated with aphidicolin (125 μM), grown at room temperature, and Abs₆₀₀ changes recorded over time.

Figure 4. Quantitatively capturing growth kinetics of compound-treated wild type and mcm2DENQ mutant strains.

A. Parameters describing cell expansion illustrated by a simulated bacterial growth curve. Growth is captured by lag time (λ), maximal growth rate (μ), and maximally achievable cell density (A). Y-axis (ln (N/N0)) shows the natural logarithm of cell expansion over initial seeding density (t=0). B. Actual growth curves of the Δ3xMDT WT and Δ3xMDT 2DENQ mutant strains in the presence of aphidicolin (125 μM; mean ± S.D; n=192). C. Quantification of magnitude of separation between wild type (WT, blue symbols) and mutant (2DENQ; yellow symbols) by Z’ factors and S/B derived from single growth curve parameters A, μ,and λ. D. Separation of the two classes by combining the three Gompertz parameters visually (3D plot, left) and numerically after LDA (right).

We used the resulting growth curves to derive λ, μ, and A. Visually, the data showed that aphidicolin both slowed the expansion rate of the 3xMDT 2DENQ strain relative to the wild type strain and prevented cultures from reaching maximum density (Figure 4 B). Surprisingly, using the Z’ criterion ³⁰, all three calculated parameters approached an acceptable screening threshold (Figure 4 C). The best separation between the wild type and mcm2DENQ strains was obtained with the slope parameter μ (Z’ = 0.78). However, the signal to background ratio (S/B) was small (< 2-fold), indicating that basing our analysis completely upon μ would make the assay susceptible to changes in well-to-well variability ³⁰. Similarly, both λ and A had shortcomings: Z’-factors were below 0.5, and the S/B ratio for A was very small (Figure 4 C).

While each single parameter provided some degree of separation between wild type and 2DENQ strains, separation was visually improved when plotting data in three dimensions (Figure 4D). We thus reasoned that combining the three Gompertz parameters into a single criterion would improve both the screening statistics and our ability to discriminate phenotypically diverse “hits”. Previous attempts to improve assay performance in systems with high inherent variability (i.e. transgenic zebrafish³³ or cell-based high-content analysis^{31, 32, 34}) used LDA²⁴, a multivariate method that maximizes separation of data populations into two predefined classes ²⁴ (e.g., wild type vs. mcm2DENQ cells). We therefore applied LDA to the three parameters from the Gompertz equation and found that LDA considerably improved the statistical sensitivity of our analysis as judged by an improved Z’ over the most robust single parameter (μ) and a visually increased separation between the means of WT and 2DENQ growth compared to any single parameter (Figure 4 D).

Extending the range of detectable growth inhibition with LDA.

To examine how well our LDA analysis platform works with less robust inhibition, we extended our analysis using three different concentrations of XL413. At low concentrations (30 μM), the rate of expansion provided the most informative difference between the two test strains, however both strains recovered to a nearly similar final density (Figure 5 A). Consequently, the slope parameter with this drug concentration robustly distinguished wild type from mcm2DENQ (Z’ = 0.72) whereas lag time and final density did not (Z’ < 0) (Figure 5 B). However, at an XL413 concentration of 100 μM, both rate of expansion and final density were significantly different, and each single parameter approached or surpassed HTS criteria. Finally, at the highest concentration of XL413 tested (300 μM) assay performance declined because of increased toxicity to the wild type strain; as a result, no single parameter achieved HTS performance (Figure 5 B).

Figure. 5. LDA extends the range of detectable growth inhibitory phenotypes.

A. Growth curves of Δ3xMDT WT (blue) and Δ3xMDT 2DENQ (red) in the presence of various concentrations of XL413 illustrate three different growth inhibitory phenotypes. Each data point is the average of 192 replicates ± S.D. B. Quantitative characterization of data in (A) by modified Gompertz analysis. C. Histograms of individual parameters vs. LDA analysis illustrate magnitude of separation of positive and negative controls and improvement by LDA. Note that the Y axis in panel (A) shows the natural logarithm of cell expansion (ln(N/N0)) whereas the table in (B) depicts non log-transformed data to derive statistics. ^a non log-transformed expansion; ^b No S/B value is reported because of the nature of the LDA values (see Materials and Methods for details).

We next applied LDA to these results and found that it substantially improved the ability to quantify XL413 growth inhibition. At the highest XL413 concentration (300 μM), where no single growth parameter was useful for screening purposes (Z’ ranging from −0.3 to 0.07), LDA elevated the assay into screenable range (Z’ = 0.41) ³⁰. At the intermediate XL413 concentration (100 μM), LDA considerably improved Z’ (0.57 to 0.72). At the lowest concentration, where one single parameter (μ) already provided robust separation of wild type vs. 2DENQ growth, LDA led to a minor improvement in Z’ (0.80 compared with 0.78 using μ alone). These results can be readily visualized using histograms of well populations, which illustrate the magnitude of separation between the two strains for the four parameters at the various concentrations of XL413, with LDA being superior to even the best single parameter (Figure 5 C).

Discussion

Our results support the feasibility of using a cell-based approach and an informative mcm allele to identify novel DNA replication inhibitors; our pilot screen has identified previously unknown inhibitors of the mcm2DENQ mutant, and these inhibitors have led to the development and statistical validation of a novel HTS. As discussed further below, these results support our premise that such screens will lead to the identification of inhibitors that specifically target cells engaged in pathological forms of DNA replication. Such inhibitors would be expected to be especially useful in the treatment of various proliferative diseases such as cancer.

Using synthetic growth defects to isolate DNA replication inhibitors.

We found that cells carrying the mcm2DENQ allele were differentially sensitive to an inhibitor of DNA polymerase alpha (aphidicolin ⁴¹) and the DNA-dependent replication kinase CDC7/DBF4 ⁵ (Figure 3). Recently, a synthetic growth defect was observed between the mcm2–1 and cdc7–1 alleles ¹⁶, an observation that independently supports and validates our observations with XL413 (an inhibitor of CDC7). In contrast, our observation of aphidicolin sensitivity has potentially identified a new synthetic interaction between DNA polymerase alpha and the mcm2DENQ allele. Both compounds primarily inhibit the enzymatic activities of these replication factors rather than functioning as DNA damage agents. The results suggest that our screen can identify enzymatic inhibitors rather than compounds that cause apparent synthetic growth defects by generating DNA damage.

The targets identified in our pilot screen (CDC7 and DNA polymerase alpha) may represent the proverbial tip of the iceberg. A recent high-throughput genetic interaction study utilizing the mcm2–1 allele both broadly confirms our observed mcm2DENQ synthetic growth defects (Figure 1), but additionally demonstrates that the mcm2–1 mutant has synthetic growth interactions with several hundred additional genes (¹⁶, http://thecellmap.org/tabular/?n=876). The strongest among these interactions include conditional or hypomorphic alleles of a large number of replication initiation factors that are known to physically interact (e.g., alleles of ORC, CDC45, Mcm2–7, CDC7/DBF4, GINS complex, CDC6, and CDT1). Thus, future inhibitors identified using our mcm2DENQ HTS might be expected to specifically target a broad range of replication initiation factors.

Overcoming challenges in synthetic growth inhibition HTSs.

Phenotypic HTS for compounds that cause synthetic growth defects requires measuring differential growth inhibition in two matched cell lines. This differential can vary substantially depending on the biological system under investigation, leading to small and variable assay windows, which are undesirable for large scale HTS ³⁰. Therefore, a screening paradigm is often chosen that measures growth inhibition of the more sensitive cell line in the presence and absence of a putative inhibitor. “Hits” that emerge from such a screen are then rescreened in both the wild type and test strains at a variety of concentrations to obtain IC₅₀ values; these values are subsequently used to eliminate compounds that cause non-specific growth inhibition from those that are truly specific for the initial screening goal. This strategy maximizes Z’-factors because it is easy to distinguish growth of treated and untreated cells, but has scientific and logistic shortcomings. First, inhibition of growth of a single cell line is not a relevant positive control as it is the differential between wild type and mutant cells that defines synthetic lethality. Second, this paradigm results in a large number of false positives as not all compounds that inhibit the mutant strain will also show differential growth inhibition, and these will only be detected in secondary assays. These secondary screens are costly, both in terms of time and the expense of purchasing fresh samples of a large number of non-selective inhibitors. The data from our small pre-screen with DNA damaging agents show that of the fifteen agents assayed, eight inhibited the growth to the 2DENQ mutant, but only three showed differential inhibition of WT vs. mutant, a false positive rate of 63% (Figure 2).

Towards reducing the burden of extensive secondary screens to eliminate large numbers of uninformative compounds, we designed our assay to directly compare growth inhibition of the test strain (e.g., containing the mcm2DENQ mutation) with the control parental strain (e.g., wild type). However, although this criterion reduces the number of false positives, it also narrows the usable screening window of the assay, as the growth of both strains will likely be inhibited to at least some extent. False negatives could also be a problem because HTS is usually limited to a single concentration of compounds and growth inhibition is measured at a fixed endpoint, potentially missing agents with desirable characteristics that may not show optimal differential activity at the concentrations and times chosen for the screen.

To overcome both of these obstacles, we acquired full growth curves and performed LDA ²⁴ to combine the three kinetic parameters A, μ, and λ, which collectively describe the growth kinetics of single celled-organisms (Figure 4A), into a single parameter, the LDA value. As expected from prior work by others ^{31, 32, 34} and our recent zebrafish HTS ³³, using LDA values to calculate Z’-factors substantially improved assay performance compared with each single parameter (Figures 4 and 5). As with all types of multivariate analysis, the magnitude of improvement was most pronounced under conditions where the assay was more variable. When one parameter by itself already robustly separated populations, additional parameters did not contribute substantially more. Where the two classes are poorly separated, however, the linear combination can have dramatic effects. We have observed this in our zebrafish HCA, where high inherent variability and limiting sample sizes hindered HTS performance, but LDA resulted in a screenable assay ³³. In the experiments with XL413, where at the highest concentration assay performance declined substantially due to inhibition of the wild type cells, LDA also moved the assay into screenable range (Figure 5). Therefore LDA should help stabilize Z’-factors by buffering sources of assay variability, such as those contributed by substantial or inconstant baseline inhibition or environmental conditions (e.g., a stack of plates kept at ambient temperature, which can fluctuate significantly).

Much more important than an increase of HTS assay robustness, however, is the fact that LDA does not just increase Z’ factors, which are an accepted measure of assay robustness, but are based on positive controls that may or may not reflect the range of compound behaviors in a screening library. The biggest advantage of LDA is its ability to capture growth inhibitory phenotypes not detected by any single parameter alone (Table 1). For example, the maximum expansion parameter might detect conditions where growth inhibition is sustained and results in lower cell densities at the end of the experiment. The slope parameter might detect sustained growth inhibition, but not necessarily delayed growth if the culture fully recovers over time. Measurements of lag time are based on the linear portion of the growth curve during log phase. They are therefore sensitive to changes in slope, and might even be falsely detected as smaller than control if the slope is shallow; in fact the dependence of lag time measurements on the shape of the growth curve is probably the reason why in our experiments λ was consistently the lowest performing parameter. Of the four parameters, only LDA can capture all growth inhibitory phenotypes regardless of growth kinetics. From a screening perspective, this feature of LDA is especially important as in HTS there is no a priori knowledge of a compound’s potency and temporal characteristics, and compounds that do not show optimal synthetic growth inhibition at the single concentration or time point typically used in HTS would be missed. Thus LDA is expected to extend the detectable range of synthetic growth inhibitory phenotypes, leading to a more efficient exploration of chemical space compared with single parameters. Whether LDA will prove to be the best analysis tool to detect positives in compound screens remains to be seen; to the best of our knowledge the method has been applied only to assay development because it maximizes the difference between known classes (MIN and MAX), but it is not clear how well diverse compounds will conform to the criteria set by the positive and negative controls. In this regard, principal component analysis (PCA) might be a viable alternative as it does not include the group information. Once compound screening data are available, it will be interesting to compare alternative multivariate methods such as PCA or the kernel-based discriminant analysis for data sets with non-linear structure ³⁴.

Table 1. Possible growth inhibition phenotypes and ability to detect them during HTS.

Table 1 provides a summary of possible growth inhibitory phenotypes; while single parameters can occasionally detect positives, only LDA captures all possible growth inhibitory phenotypes, which should permit discovery of a much wider range of agents and reduce the rate of false negatives. Please refer to the main text for more detail.

Growth Inhibition Phenotype	Expected Effect on Growth Curve Parameters			Synthetic Growth Defect captured by
	A (max. expansion)	λ (lag)	μ (slope)	A alone	λ alone	μ alone	LDA
Delayed growth with recovery over time	Same as control; time dependent	Increased	Similar or smaller	No	Yes	Yes / No	Yes
Delayed growth with sustained inhibition	Lower than control	Smaller or not measurable	Smaller	Yes	Yes / No	Yes	Yes
Acute toxicity with recovery over time	Same as control; time dependent	Increased	Similar or smaller	No	Yes	Yes /No	Yes
Acute toxicity with sustained kill	Lower than control	Smaller or not measurable	Smaller	Yes	Yes / No	Yes	Yes

In summary, we have developed a robust, HTS compatible assay and analysis platform to identify compounds that cause synthetic growth defects in yeast. Because the analysis is applicable to any assay measuring cell growth, it has the potential to significantly impact not only the field of DNA replication, but also the discovery of antibacterials, antifungals, and of novel targets in various biologies through chemical probe development. The two agents that emerged from our limited pre-screen already suggest two novel, previously unappreciated modulators of Mcm2–7 activity, and we posit that HTS is bound to uncover many more.

Abbreviations:

DMSO

dimethyl sulfoxide

5-FOA

5-fluoro-orotic acid

HCA

high-content analysis

HTS

high-throughput screening

HU

hydroxyurea

LDA

linear discriminant analysis

MDT

multidrug transporter

PCA

principal component analysis

PCR

polymerase chain reaction

WT

wild type

YPD

yeast peptone dextrose media