For scientific reasons and because of the needs of the pharmaceutical and agrochemical industries, there is much current interest in detecting the site or mode of interaction between an exogenous ligand and a cell or organism, especially in identifying new ones (6
). Such studies are nowadays typically carried out by using high-throughput methods, but there is often an inverse relation between the speed of an assay (involving, e.g., a cloned receptor with a fluorescence readout) and the amount of information it contains (which in this case is restricted to the target of interest). However, array-based methods are showing promise in this regard (3
). Genome-wide screens can also provide such information (9
) but require numerous strains or cell lines to be studied in parallel. Methods with high information content would combine the virtues of screening just a small number of strains with a high-dimensional readout similar to that provided by the array-based methods.
Metabolic control analysis (see, e.g., references 7
, and 29
) tells us that while changes in the activity of a target enzyme tend to have only small effects on the flux through metabolic pathways, they can and do have substantial effects on the concentrations of metabolic intermediates. Since the use of metabolomics, especially in the form of “metabolic fingerprinting,” has a much higher throughput and is much cheaper than are, say, transcriptomics and proteomics (8
), it makes an attractive candidate for mode-of-action studies (18
) and has been used to advantage by Ott and colleagues (2
), where the metabolic fingerprints of cells or tissues exposed to specific substances with known targets were analyzed for their modes of binding or action by using pattern recognition techniques.
Normally, metabolomics measures (or seeks to measure) the concentrations of all the small molecules within a cell or tissue; and for purposes of functional genomics, we and others have exploited such strategies in the analysis of gene deletion mutants in Saccharomyces cerevisiae
). However, because of the need to separate cells from medium and to extract the metabolome of the former separately, these methods are not as convenient as one would wish. In consequence, we have recently shown that the analysis of the metabolic footprint, the cocktail of metabolites left behind or secreted into the medium, differs reproducibly in single-gene mutants and can be measured rapidly and conveniently via direct-injection mass spectrometry (1
) or Fourier transform infrared spectroscopy (22
The question of whether metabolic footprinting of strains treated with different pharmaceuticals might also be used in discriminating the modes of action of different substances then arises. The purpose of the present study was to explore this question, and we show here, using metabolic footprinting and both unsupervised and supervised methods of machine learning analysis, that we can indeed classify a number of fungicides according to their modes of action.
MATERIALS AND METHODS
Fungicide screening. (i) Yeast strain.
The BY4743 diploid strain (MATa/MATα ura3Δ0/ura3Δ0 leu2Δ0/leu2Δ0 his3Δ1/his3Δ1 +/lys2Δ0 met15Δ0/+) was used throughout.
(ii) Screening methodology.
A library of agrochemical fungicides supplied by Aventis CropScience was screened for activity against yeast in a Bioscreen microtiter plate (Honeycomb II plates; Labsystems) format. Stocks (25 mg ml−1) of each compound were prepared in dimethyl sulfoxide (DMSO). From these stocks, three additional stocks (2.5 mg ml−1, 250 μg ml−1, and 25 μg ml−1) were prepared by serial dilution, such that each compound could be screened at log concentrations so that a broad picture of its inhibition behavior could be obtained. Stocks were aliquoted into glass vials and stored at −40°C.
From a washed starter culture, a stock of 5 × 105 cells ml−1 was prepared, and 550 μl was aliquoted into a series of 1.5-ml sterile microcentrifuge tubes. A total of 5.5 μl of fungicide at each concentration to be tested (250, 25, 2.5, and 0.25 μg ml−1) was then added to the appropriate microcentrifuge tube and vortexed to ensure good mixing. DMSO control tubes (5.5 μl of DMSO added to 550 μl of cell suspension) were also set up. From each microtube, 100 μl was pipetted into each of five adjacent Honeycomb plate wells to produce a set of five replicate wells per treatment. Plates were then placed in the Bioscreen, where they were incubated at 30°C and shaken at medium intensity. Optical density (OD) at 600 nm was measured every 20 min for the duration of the screen (typically, 24 to 36 h).
The percent inhibition for each compound was calculated as the difference in the OD of treated cultures between 0 and 17 h expressed as a percentage of the change in the OD of the DMSO controls over the same time span.
(iii) Metabolic footprinting of fungicide-treated cultures.
All footprinting of fungicide-treated cultures was performed in (Bioscreen) microtiter plate format. For all footprinting studies of fungicide-treated cultures, each compound (which had been demonstrated to have reasonable activity against yeast following screening) was added at four different concentrations which had been found to inhibit yeast growth over the approximate range of 20 to 80% (compared with DMSO controls). The aim was to obtain an approximate fixed percentage of growth inhibition such that any changes in the footprint could be ascribed to the mode of action of the inhibitor and not, say, changes in growth rate.
Footprints were typically collected at 24 h (and no later than 28 h) at a point at which the majority of cultures appeared to have entered stationary phase (1
). Plates were then removed from the Bioscreen and 90 μl was removed from each well, placed in a sterile 0.5-ml microcentrifuge tube, and spun for 5 min. Approximately 80 μl of the resulting supernatant was then removed and stored at −40°C in polystyrene microtiter plate (Nalgene) format until mass spectrometric analysis. Samples were diluted for mass spectrometry and analyzed exactly as described previously for deletion mutant footprints (1
(iv) Mass spectrometry.
Direct-injection mass spectrometry was performed exactly as described previously (1
(iv) Numerical methods.
Principal-components (PC) analysis (PCA), discriminant function (DF) analysis (DFA), and hierarchical cluster analysis (HCA) were carried out by using programs written in MATLAB (Mathworks, Cambridge, United Kingdom) exactly as described previously (1
). Genetic programming was performed with the program gmax-bio (Aber Genomic Computing [http://www.abergc.com
RESULTS AND DISCUSSION
A typical set of growth curves of the diploid strain in the Bioscreen, at different concentrations of inhibitors, is given in Fig. 1
. It is clear that the Bioscreen provides an accurate and convenient system for registering the growth of these cells and its inhibition by the inhibitors. It also shows that, with the possible exception of fenpropimorph, there is no regrowth in the presence of the inhibitors, i.e., a phenotypic (or conceivably genotypic) emergence of resistance in the cells after they have been exposed to the inhibitors (as has been noticed, for example, in bacterial experiments [40
The effect of the different concentration categories on the inhibition profiles of the nonrespiratory inhibitors is given in Fig. 2
, and the equivalent concentrations are shown in Table 1
. The two concentrations of respiratory inhibitors used are given in Table 2
Metabolic footprints were taken from these cells and subjected to mass spectrometric analysis exactly as described previously (1
Following PCA (data not shown), DFA was performed on the data, a cross-validation analysis demonstrating that PCs 1 to 20 (99.9% of the variance) were sufficient to give reasonable clustering according to compound class without compromising the overlap of training and test data (Fig. 3
). HCA was additionally performed on the DFA scores (score, 1 to 3; averaged according to compound) in order that grouping relationships might be better visualized (Fig. 4
). Clearly, the greatest separation is between respiratory and nonrespiratory inhibitors, which form two separate branches in the resulting dendrogram. Interestingly, within the respiratory inhibitor branch, the complex III strobilurin-type inhibitors form their own subbranch, while the complex II inhibitor carboxin groups with the DMSO control form a separate subbranch. A particularly interesting feature is that the classification works over a range of doses. This robustness to the actual dose represents a significant benefit in setting up experiments, since this could save a lot of time on preliminary, dose-setting experiments. It is noticeable that the DMSO control appears on the same branch with inhibitory compounds rather than, as one might possibly hope, forming a separate branch on its own. However, even at a concentration of 1% (128 mM), the level of DMSO is somewhat above that (50 mM) which was demonstrated by Panek et al. (39
) to affect various cellular processes, and so it cannot be considered entirely benign.
In a similar vein, we trained a genetic program (4
) to evolve rules that would discriminate respiratory from nonrespiratory inhibitors. DMSO samples were removed, as their status would of course be known. To guard against overtraining, the data were split into three partitions, in accordance with accepted modelling practice (e.g., see reference 43
). The genetic programming was trained on the first partition, terminating evolution after 20 generations. This cycle was repeated 10 times. Each of the resulting 10 rules gave perfect classification on the training partition. Each rule was then applied to the second partition in order to test its generalization capability, and of the 10 rules, 6 also gave perfect classification on this partition. These six rules were then applied to the third partition (which had not been used in any way in forming or selecting the rules), and in each case a perfect classification was obtained. In partitioning the data, replicates of individual inhibitors had been retained within single partitions, and thus, the third partition contained only samples treated with inhibitors that were not represented in either of the first two partitions but whose modes of action were nevertheless correctly classified.
Each of the six successful rules used a combination of just three m/z
ratios to discriminate the classes. All but one of the six rules contained an m/z
ratio of 144, but no other m/z
ratio appeared more than once. Plotting m/z
144 in turn against each of the other variables used by the rules revealed three cases (m/z
141, 179, and 189) where clusters visually separable with straight lines were obtained (see the example of m/z
144 versus 141 in Fig. 5
). Thus, as in previously described related strategies (1
), it seems that these classifications can be performed effectively by using only a very small number of the variables from these high-dimensional data sets, and there are also likely to be other combinations of peaks in these spectra that are equally able to discriminate classes. It is possible that a chemical interpretation of these peaks may lead to the identification of useful markers for understanding the biochemical basis of these discriminations.
Within the nonrespiratory branch, it was pleasing to observe that the sterol biosynthesis inhibitors formed one branch (albeit the morpholine fenpropimorph forms a subbranch with the azole epoxiconazole) while the amino acid biosynthesis inhibitor and uncoupler form another. The separation of the respiratory inhibitors from the remaining compounds is scarcely surprising given that the former inhibitors give rise to only limited or no inhibition of growth rate. Consequently, in order to determine whether similar grouping relationships might be observed between the remaining (nonrespiratory) compounds in the absence of variance contributed by the respiratory inhibitors, data for these latter compounds were removed from the data set, and the remaining compounds were then resubjected to PCA, DFA (PC scores of 1 to 20, employing the same compound-based class system), and HCA (DFs 1 to 3, averaged according to compound). The resulting dendrogram (Fig. 6
) illustrates that the grouping relationships observed in Fig. 4
are preserved among these nonrespiratory inhibitors, in that the sterols form a group away from the rest of the inhibitors and the uncoupler is separated from the acetolactate synthase inhibitor.
The attractions of the metabolic footprinting approach are its rapidity, ease of implementation, and reproducibility, while as a pattern recognition strategy (8
), it is not designed (immediately) to determine the molecules that are most important to or responsible for any particular discrimination. However, the simplicity of some of the rules determined by the genetic programming approach means that, in many cases, just a few rules are sufficient (e.g., as shown in Fig. 5
). This result means that if a tandem mass spectrometer is available, only those masses need to be studied to achieve identification (and these could then be confirmed by other methods such as gas chromatography-mass spectrometry or nuclear magnetic resonance. Such an iterative strategy is entirely appropriate for postgenomic studies (28
In conclusion, the present work has shown, complementarily to previous studies in which gene knockouts were used (1
), that the metabolic footprints of wild-type strains treated with inhibitors with different modes of action contain sufficient information to allow such modes of action to be discriminated from the metabolic footprints alone.