Matthew F.
Brown
*a,
Rishi R.
Gupta
*b,
Max
Kuhn
c,
Mark E.
Flanagan
a and
Mark
Mitton-Fry
a
aWorldwide Medicinal Chemistry, Pfizer Global Research and Development, Eastern Point Road, Groton, CT 06340, USA. E-mail: matthew.f.brown@pfizer.com; Fax: +1 860 715 4693; Tel: +1 860 441 3522
bComputational Chemistry, Pfizer Global Research and Development, Eastern Point Road, Groton, CT 06340, USA
cNon-Clinical Statistics, Pfizer Global Research and Development, Eastern Point Road, Groton, CT 06340, USA
First published on 10th June 2011
In this work we present a number of statistical and visualization methods derived from MIC90 data designed to aid decision-making in antibacterial drug discovery research. A statistical method known as bootstrapping was applied to MIC90 raw data to uncover data trends and a metric termed Net Percent Superior (NPS) was developed to capture a strain-by-strain analysis of analogs to enable rank-ordering of similar compounds. We also present novel methods of reporting the data using a variety of visualization techniques. Furthermore, the work was cross-validated using experimental results generated with siderophore-conjugated monocarbam analogs to demonstrate the effectiveness of the various parameters and visualization techniques. The methods reported herein have been incorporated in a Scitegic Pipeline Pilot protocol to enable facile, automated generation of MIC90 analyses from experimental raw data to aid prospective medicinal chemistry design as well as retrospective analyses.
Compound | R 1 | R 2 | MIC90a (n = 51) | BMIC90b | BMIC90 95% CIc | NPSd | NPS 95% CIe | GMf | GM 95% CIg |
---|---|---|---|---|---|---|---|---|---|
a The drug concentration which inhibits the visible growth of ≥90% of the bacterial strains within a test population. b Bootstrap MIC90. c 95% confidence interval for BMIC90. d Net Percent Superior. e 95% confidence interval for BMIC90. f Geometric mean. g 95% confidence interval for geometric mean. h Published MIC90 data (n = 206).6b | |||||||||
1 (MC-1) | –H |
![]() |
1 | 0.95 | (0.5, 8) | 0 | — | 0.27 | (−0.45, 1.00) |
2 | –Me |
![]() |
0.5 | 0.52 | (0.5, 0.5) | 70.6 | (55, 86) | 0.14 | (−0.03, 0.31) |
3 | –Me |
![]() |
0.5 | 0.50 | (0.25, 1) | 19.6 | (2, 37) | 0.21 | (0.06, 0.37) |
4 | –Me |
![]() |
1 | 0.79 | (0.5, 2) | 2 | (−16, 20) | 0.27 | (−0.35, 0.89) |
5 | –Me | –Me | 1 | 1.11 | (1, 4) | −54.9 | (−73, −35) | 0.39 | (−0.05, 0.82) |
BAL30072 | — | — | 4 (8)h | 3.02 | (1, 8) | −82.4 | (−92, −71) | 0.86 | (0.14, 1.58) |
COMPOUND LINKS Read more about this on ChemSpider Download mol file of compoundMeropenem |
— | — | >64 | 59.5 | (32, 64) | −72.5 | (−90, −53) | 2.85 | (−3.75, 9.44) |
The “Standard Heatmap” view of these data was created (Table 2) wherein the activity of each compound vs. each strain is color-coded according to a rough estimate of the activity required to kill the pathogen in a clinical setting (MIC ≤ 0.25: dark green, MIC = 0.5–1: light green, MIC = 2: yellow, MIC = 4–8: orange and MIC ≥ 16: red). The table is sorted left to right based on the MIC90 value with most potent compounds appearing on the left. The table is also sorted vertically based on the Meropenem data. From this view, one can immediately appreciate the exceptional level of potency the monocarbam analogs generally achieve vs. these clinically relevant strains. It also highlights those few strains which demonstrate a moderate level of resistance to the chemotype (1766-00, PA-1757, 1509-08, 1510-08) and the effectiveness of the monocarbam analogs vs. the competitor agent BAL30072. This view also illustrates the superior performance of the monocarbam analogs vs. a range of strains highly resistant to COMPOUND LINKS
Read more about this on ChemSpider
Download mol file of compoundMeropenem. Bacteria have developed effective mechanisms of resistance to many β-lactam drugs, including expression of a wide variety of β-lactamases, expression of efflux pumps, downregulation of porin channels and active site mutations of the penicillin binding proteins, which are the biological targets of β-lactam antibacterial agents.8 One can appreciate the additional value this kind of heatmap view could provide if the bacterial strains utilized in an MIC90 study were characterized with regard to their resistance phenotypes as this would enable one to potentially correlate potency (or the lack thereof) with the presence of one or more mechanisms of resistance. This heatmap view could also be useful when conducting retrospective analyses on large datasets.
![]() |
One limitation of the standard heatmap view is that it does not effectively highlight relative potency differences between similar compounds and, therefore, does not enhance the ability to rank-order closely related analogs, especially those with equivalent MIC90 values. One needs to appreciate that MIC90 experiments are rather labor intensive, therefore, decisions around compound advancement are often driven by learnings derived from a single experiment. Given the limited sample size of an intermediate MIC90 panel like this, coupled with the inherent error associated with individual MIC values (approximately +/−2-fold), decisions based on these data can be somewhat speculative. Oftentimes, inspection of MIC90 raw data suggests the potential for the MIC90 value to change if the experiments were repeated. This tendency towards moving upward or downward can be important to recognize and utilize in decision-making and can be visualized in Table 3. Here, the MIC data in each column have been independently sorted lowest to highest value and the row representing the MIC90 value is highlighted in purple. The orange shaded area is provided to illustrate whether the MIC90 value is “trending” toward a higher or lower value. Ideally, the MIC90 value would fall roughly in the middle of the orange shaded area as this would provide confidence that if the experiments were repeated, minor changes in activity throughout the panel would not lead to a change in the MIC90 value. The closer the MIC90 value is to the upper or lower boundary of the orange shaded area, the greater the likelihood that the MIC90 could change if the experiments were repeated. Based on this discussion, one might surmise that compounds 1, 4 and BAL30072 are leaning toward a lower MIC90 value, compound 2 is leaning toward a higher value and the MIC90s for compounds 3, 5 and COMPOUND LINKS
Read more about this on ChemSpider
Download mol file of compoundMeropenem appear fairly solid. Medicinal chemists conducting antibacterial research are accustomed to looking for these tendencies, and the ability to quantify them as numerical values would be useful.
To capture this MIC90 tendency, we introduce a new parameter based on the use of a statistical methodology called the bootstrap.9 In general, the bootstrap method is a re-sampling technique that uses many alternate versions of the data, called bootstrap samples, to estimate the statistical distribution of the MIC90. A bootstrap sample is a sample taken with replacement, meaning that after a species is selected to go into the bootstrap sample, it can be selected again. The size of the bootstrap sample is the same as the original dataset (51 data points in this case), and will likely contain duplicate values of some strains due to the re-sampling process. A large number of bootstrap samples are taken (1,000,000 times for this analysis) and, for each, the MIC90 value is computed. These computed MIC90s constitute the “bootstrap distribution” for the statistic. Statistical theory allows the use of this distribution to estimate statistical summaries of the MIC90, such as the average or confidence intervals. The bootstrap estimate of the MIC90, denoted as the BMIC90 (Fig. 1), is the simple average from the bootstrap distribution. Since the MIC90 is defined to be a dilution within a series, the bootstrap method for defining confidence intervals results in highly conservative (i.e. wide) intervals and is not useful with regard to rank-ordering compounds (Table 1). However, other procedures can be used to determine if there are statistically meaningful differences in MIC90 values between compounds. A permutation test9 was used to estimate the p-value for the hypothesis that the MIC90 values are the same for two compounds. If the p-value is small, this would indicate that there is evidence that the values are not equal. However, when testing all combinations of compounds, there is a concern that multiple statistical tests will overly inflate the confidence intervals. To compensate for this, the raw p-values were adjusted to control the false positive rate (also known as the false discovery rate or FDR).10 While this procedure is more conservative than no adjustment, it is more liberal than classical p-value correction procedures that protect against any false positive results. A corrected p-value less than 0.05 suggests that a statistically meaningful difference in MIC90 value exists. Table 4 shows the FDR values for each pair-wise comparison and suggests that the MIC90 values for all compounds in this set are statistically different, with the exception of compound 1 being indistinguishable from compound 4. Strictly speaking, the FDR values alone do not enable rank-ordering of analogs, but can be used in conjunction with other parameters as will be discussed later.
![]() | ||
Fig. 1 Bootstrap method. |
FDR (adjusted p-values from pair-wise analysis) for MIC90 values | |||||||
---|---|---|---|---|---|---|---|
1 | 2 | 3 | 4 | 5 | BAL30072 |
COMPOUND LINKS Read more about this on ChemSpider Download mol file of compoundMeropenem |
|
a 1 and 4 are not different based on p-value. | |||||||
1 | — | 0.0000 | 0.0414 | 0.4917a | 0.0002 | 0.0000 | 0.0000 |
2 | — | 0.0035 | 0.0000 | 0.0000 | 0.0000 | 0.0000 | |
3 | — | 0.0414 | 0.0000 | 0.0000 | 0.0000 | ||
4 | — | 0.0004 | 0.0000 | 0.0000 | |||
5 | — | 0.0000 | 0.0000 | ||||
BAL30072 | — | 0.0001 |
The “Finland-o-gram” is occasionally utilized to visualize the differences between similarly active compounds.11 This involves plotting the cumulative % inhibition of susceptible strains vs. the MIC values (Fig. 2). The curves for the most potent analogs are left-shifted relative to less potent analogs. To further aid in visualizing differences between closely related analogs the “Relative Heatmap” view was designed (Table 5). In this view, the activity of a set of new analogs is compared to the activity of a related reference compound in a strain-by-strain analysis. The color in each cell is determined based on how well the new compound performed vs. a particular bacterial strain relative to the activity of the reference compound (1 in this example).12 The compounds in the table are arranged left to right based on MIC90 potency and the data are sorted vertically based on the individual MIC values for the reference compound. In this view, it is readily apparent that most of the compounds are not superior to the reference compound 1 with compounds 2 and 3 being the possible exceptions. This is not surprising given that the MIC90 values for 2 and 3 are superior to 1 (0.5 μg mL−1 and 0.5 μg mL−1vs. 1 μg mL−1). This view clearly demonstrates that the superiority of 2 relative to 1 is consistently observed across the entire panel of strains, including the toughest to treat strains which are least susceptible to 1 (magenta shaded strains in Table 5). The superiority of 3 to 1 is somewhat less consistent over the entire panel, but again is demonstrated consistently vs. those strains least susceptible to reference compound 1. A more interesting comparison would be compounds 1 and 5 as these have equivalent MIC90 values (1 μg mL−1), yet the Relative Heatmap view clearly suggests that 5 is inferior to 1. Just as the BMIC90 parameter was developed to capture MIC90 data trends, a new parameter termed the “Net Percent Superior” or NPS was developed to capture trends observed in the Relative Heatmap view. This parameter is calculated by subtracting the percentage of strains with activity inferior to the reference compound from the percentage of strains with activity superior to the reference compound. The reference compound will obviously have an NPS value equal to 0; analogs with NPS values >0 are defined as potentially being superior to the reference compound and those with an NPS value <0 are potentially inferior. Since these values have higher resolution than the MIC90 values, the bootstrap procedure can be used here to generate meaningful confidence intervals (Table 1 and Fig. 3). The weakness of this parameter as calculated is that it does not account for the magnitude of activity differences between the reference compound and the analogs; a 2-fold difference in activity is treated the same as a 16-fold difference. As a result, this parameter does not effectively rank-order compounds that consistently demonstrate activity different from the reference compound on a strain-by-strain basis. For example, while compound 5 is clearly superior to COMPOUND LINKS
Read more about this on ChemSpider
Download mol file of compoundMeropenem based on MIC90 values (1 μg mL−1vs. >64 μg mL−1), the NPS values (−54.9 vs. −72.5) and the corresponding confidence intervals (Fig. 3) do not reflect this superiority. That said, the goal of this work was to develop methods to identify and rank-order the most promising analogs in a series with similar activity. The NPS parameter clearly identifies compounds 1–4 as being superior to 5, BAL30072 and COMPOUND LINKS
Read more about this on ChemSpider
Download mol file of compoundMeropenem and can contribute to rank-ordering as discussed below.
![]() | ||
Fig. 2 Finland-o-gram. |
![]() |
![]() | ||
Fig. 3 95% confidence intervals for Net Percent Superior (NPS) values. |
There are a number of examples in the literature describing statistical analyses of MIC90 raw data (arithmetic mean, geometric mean, median, mode, etc.) to enable rank-ordering compounds with the use of the geometric mean being somewhat common.13 Therefore, we have also included the geometric mean in our analysis for comparison (Table 1). As was the case with the BMIC90 parameter, confidence intervals for the geometric mean values are also quite wide and not useful with regard to rank-ordering analogs. As alluded to above, antibacterial research teams often make decisions around compound progression based on MIC90 data. Rank-ordering this set of analogs based solely on the MIC90 data would provide the following: 2 = 3 > 1 = 4 = 5 > BAL30072 > COMPOUND LINKS
Read more about this on ChemSpider
Download mol file of compoundMeropenem. By using the parameters described above, we can attempt to provide a greater degree of separation. Recall that the FDR values strongly suggest that MIC90 values for all compounds are statistically different with 1 and 4 being the exceptions. With this in mind, recognizing the ranking trends observed for the BMIC90, NPS and geometric mean values could lead to the following rank-order: 2 > 3 > 1 ≈ 4 > 5 > BAL30072 > COMPOUND LINKS
Read more about this on ChemSpider
Download mol file of compoundMeropenem. Obviously, given the wide confidence intervals associated with the BMIC90 and geometric mean values, this ranking cannot be stated with 100% confidence. Drug research teams regularly make decisions with regard to compound advancement based on a variety of information (potency, pharmacokinetics, safety, stability, ease of synthesis, etc.). The methods described above were designed to illustrate potential trends in potency data for single experiments; there is an inherent risk that the analysis could occasionally mislead. That said, the reality of antibacterial drug research involves teams making decisions based on MIC90 data, and while the methods described herein do not enable decisions with 100% confidence, we are hopeful that they may lead to better informed, higher quality decisions.
Footnote |
† Electronic supplementary information (ESI) available. See DOI: 10.1039/c1md00095k |
This journal is © The Royal Society of Chemistry 2011 |