M23, 4th ed.

January 2016
Replaces M23-A3
Development of In Vitro Susceptibility Testing Criteria and Quality
Control Parameters
Mair Powell, MD, FRCP, FRCPath James S. Lewis II, PharmD
Kerry Snow, MS, MT(ASCP) Ryan Owen, PhD
Linda A. Miller, PhD Elizabeth Palavecino, MD
Sujata M. Bhavnani, PharmD John H. Rex, MD, FACP
Patricia A. Bradford, PhD John D. Turnidge, MD
Sharon K. Cullen, BS, RAC Matthew A. Wikler, MD, MBA, FIDSA
Seong H. Jang, PhD
Aryun (Eileen) Kim, PharmD

Abstract
Clinical and Laboratory Standards Institute document M23—Development of In Vitro Susceptibility Testing Criteria and Quality
Control Parameters offers guidance for developing interpretive criteria and QC ranges for antimicrobial susceptibility tests against
aerobic and anaerobic bacteria, and selected fungi performed by CLSI antimicrobial susceptibility testing standards. It describes
the data used by the Subcommittees on Antimicrobial Susceptibility Testing and Antifungal Susceptibility Tests to establish these
interpretive criteria and QC ranges for antimicrobial agents, including microbiological data, pharmacokinetic and
pharmacodynamic characteristics, and clinical data. As antimicrobial agents are used in practice, additional experience accrued
may be used to reassess interpretive criteria or QC ranges. Users of these guidelines should understand that susceptibility test results
cannot predict clinical outcomes with absolute certainty. They should be used along with the best clinical judgment and laboratory
support to draw the best conclusions to serve the patient.

Clinical and Laboratory Standards Institute (CLSI). Development of In Vitro Susceptibility Testing Criteria and Quality Control
Parameters. 4th ed. CLSI guideline M23 (ISBN 1-56238-925-4 [Print]; ISBN 1-56238-926-2 [Electronic]). Clinical and Laboratory
Standards Institute, 950 West Valley Road, Suite 2500, Wayne, Pennsylvania 19087 USA, 2016.

The Clinical and Laboratory Standards Institute consensus process, which is the mechanism for moving a document through
two or more levels of review by the health care community, is an ongoing process. Users should expect revised editions of any
given document. Because rapid changes in technology may affect the procedures, methods, and protocols in a standard or
guideline, users should replace outdated editions with the current editions of CLSI documents. Current editions are listed in the
CLSI catalog and posted on our website at www.clsi.org. If you or your organization is not a member and would like to become
one, and to request a copy of the catalog, contact us at: Telephone: +1.610.688.0100; Fax: +1.610.688.0700; E-Mail:
customerservice@clsi.org; Website: www.clsi.org.

Copyright Property of CLSI for use with PK-PD FDA workshop.

Downloaded on 4/27/2017.
 M23, 4th ed.

Chapter 5: Minimal Inhibitory Concentration Interpretive Criteria

This chapter includes:

 Data used for establishing interpretive criteria

 Additional considerations for the drug sponsor
 Elements to provide in the clinical trial summary

Four categories of data are useful in establishing interpretive criteria. The typical use(s) of these data are
detailed below. Each of these data sources has distinct strengths and limitations, and the identification of
interpretive criteria depends upon consideration of all four data types, as described in Subchapter 5.5.

Category Typical Considerations

ECV  What is the typical distribution of MICs for the target pathogens?
 What is the performance of the in vitro susceptibility devices used to measure
MICs?
 Is a disk diffusion device possible? If so, how do zone diameters relate to MICs?
 How do the MICs of isolates collected in clinical trials compare to the wild-type
distribution?
 What is the ECV?
Nonclinical PK-  Based on nonclinical models, what is the PK-PD index most closely linked with
PD Cutoff antimicrobial effect?
 Based on nonclinical models, what magnitude of this PD index is needed for an
antimicrobial effect?
 What is the PK of the test antimicrobial agent in patients with the type(s) of
infections for which it is intended to be used?
 In these patients, which covariates (eg, age or renal function) influence PK? Do
dosage adjustments for these covariates permit consistent exposure for treated
patients?
 What is the nonclinical PK-PD cutoff?
CER Cutoff  What is the PK of the test antimicrobial agent in patients with the type(s) of
infections for which it is intended to be used?
 In these patients, which covariates (eg, age or renal function) influence PK? Do
dosage adjustments for these covariates permit consistent exposure for treated
patients?
 What is the correlation between MIC values and likely efficacy predicted in
patients based on CER relationships for efficacy (exposure-response
relationship) in infected patients and using human PK?
 What is the CER cutoff?
Clinical Cutoff  What outcomes were observed in clinical trials?
 Do these outcomes predict a cutoff value that is consistent with the cutoffs
predicted by the nonclinical PK-PD cutoffs and the CER cutoffs?
 Is there a correlation of patient outcomes with the pathogens isolated in clinical
trials?
 Is there a correlation of patient outcomes with the MICs of the pathogens
isolated in clinical trials (simple correlation of outcomes to MIC), ie, is it
possible to identify a clinical cutoff?
 What is the clinical cutoff?
Abbreviations: CER, clinical exposure-response; ECV, epidemiological cutoff value; MIC, minimal inhibitory concentration; PD,
pharmacodynamic; PK, pharmacokinetic; PK-PD, pharmacokinetic-pharmacodynamic.

©Clinical and Laboratory Standards Institute. All rights reserved. 31

Copyright Property of CLSI for use with PK-PD FDA workshop.
Downloaded on 4/27/2017.
 M23, 4th ed.

With respect to determination of MIC interpretive criteria, Subchapters 5.1, 5.2, 5.3, and 5.4 consider each
of these data sources in turn and indicate the types of data that are considered essential elements of
submissions to the Subcommittee on AST. Subchapter 5.5 discusses approaches to balancing the
information obtained from these different sources in the decision-making process for MIC interpretive
criteria.

Similarly, Chapter 6 considers the data needed to determine disk diffusion interpretive criteria.

5.1 Epidemiological Cutoff Value

This subchapter considers the microbiological data that should be assembled to support the selection of the
ECV as defined in Subchapter 1.3.2, and the methods for selecting ECVs.

* Mechanism of Action Studies

Sponsors should report what is known about the mechanism of action of a new antimicrobial agent. In
particular, sponsors should report study results that demonstrate the mechanism of action (eg, inhibition of
cell wall synthesis, lysis of cell membrane, protein synthesis, and inhibition of DNA or RNA replication).
These reports should provide data to substantiate both physiological and morphological effects on the
microorganisms. Understanding the mechanism of action can provide a basis for understanding the
development of resistance.

The report should include the chemical structure and a description of any structural or biological similarities
to known antimicrobial agents together with a discussion of any differences there may be in the mechanism
of action (eg, for β-lactam agents there should be a description of any differences there may be in the affinity
for penicillin-binding proteins in relevant organisms for the new and closely related existing agents).

Reports of studies evaluating microbial killing (eg, microbial kill curves) should be included along with
reports of studies on mechanism of action. A discussion of the intracellular activity of the antimicrobial
agent should also be provided where appropriate.

* Mechanism of Resistance Studies

Resistance mechanisms may limit the clinical effectiveness of an antimicrobial agent. Therefore,
characterization of the mechanisms mediating resistance and their distribution within the proposed target
pathogens is important for delineating the potential clinical usefulness of the antimicrobial agent. Common
mechanisms of resistance include alterations of the antimicrobial agent by enzymes (eg, β-lactamases or
aminoglycoside-modifying enzymes), impermeability or efflux mechanisms that prevent access of the
antimicrobial agent to target sites, and changes in the target site that result in reduced affinity of the
antimicrobial agent. Whenever possible, it is recommended that sponsors provide the genotypic
characteristics of resistance mechanisms.

For some organisms and antimicrobial agents, resistance occurs in only a proportion of the overall
population (ie, heteroresistance). Testing should be done to detect the presence of heteroresistance. In
addition, studies on the frequency of spontaneous resistance to the new agent should be provided.

When resistance mechanisms against agents of the same antimicrobial class as the antimicrobial agent under
consideration are known, a sample of organisms with MICs or zone sizes in various parts of the population
distribution should be tested for the presence of that resistance trait. The results should demonstrate that
these resistance mechanisms can be recognized by the susceptibility testing methodology.

©Clinical and Laboratory Standards Institute. All rights reserved.

32
Copyright Property of CLSI for use with PK-PD FDA workshop.
Downloaded on 4/27/2017.
 M23, 4th ed.

Antibacterial Spectrum of Activity

5.1.3.1 * Susceptibility Test Methods

 Sponsors should describe the methods used for generating susceptibility data in nonclinical studies and
in clinical trials (eg, microbroth dilution, agar dilution, disk diffusion).

 See Chapter 2 regarding establishing the reference methods for an antimicrobial agent and the
validation of alternative methods to the reference methods. For example, if sponsors use freeze-dried
panels to assess the MIC of the drug during clinical trials, they should conduct a comparative study to
demonstrate comparability of MIC results for the frozen and freeze-dried panels.

5.1.3.2 * Spectrum of Activity

Sponsors should evaluate the activity of an antimicrobial agent against a test panel of organisms that are
relevant to the intended clinical uses, including individual species of interest and species that may be
associated with group-level descriptors (eg, Enterobacteriaceae or viridans group streptococci). The
number of isolates tested and their diversity are important (eg, geographic distribution, relevant clinical
genera and species, relevant resistance mechanisms).

Dilution and disk diffusion tests should be performed on a large collection of isolates, which should include
examples of clinically relevant isolates appropriate both for the class of the agent being evaluated and for
its anticipated clinical use. The collection should include isolates showing important resistance
mechanisms, including resistance to antimicrobial therapies commonly used for indications sought by the
new agent. Refer to Appendix A for additional guidance on organism selection and sample size.

Sponsors should identify whether there are subtypes (eg, genotypes, serotypes, or multi-locus sequencing
types) that correlate with a resistance phenotype, and should include these in the test panel. Organisms that
demonstrate heteroresistance to the antimicrobial agent under study should also be included, if applicable.

The presentation of the data on the spectrum of activity should include:

 * A description of the susceptibility testing method(s) used to determine the activity of the antimicrobial
agent; if an experimental method is used, then the details of the method and its performance
characteristics in the laboratory where such testing was performed should be included.

 * A description of all susceptibility testing QC measures; all susceptibility test results should be
accompanied by a summary of QC data obtained in MIC tests performed after QC parameters were set.

 * The MICs of the antimicrobial agent for each organism (by genus- or species-level identification)
presented in tabular form to display the numbers tested, MIC range, and the MIC50 and MIC90.

 * Frequency distribution of number of isolates at each MIC in the form of histograms showing the
number of isolates (y-axis) with MICs at each concentration of the antimicrobial agent (x-axis),
spanning a range of concentrations that encompass all MICs measured or reaching the upper limit of
concentrations feasible to test. The x-axis is customarily drawn in intervals of two-fold increasing
antimicrobial concentrations (see Figure 1).

 * For each bacterial pathogen relevant to the intended clinical use, there should be additional histograms
(designed as above). These histograms should include 1) a presentation of isolates obtained in clinical
trials compared to all other isolates tested by MIC (see Figure 1), and 2) a presentation of MICs
according to the presence or absence of phenotypic resistance, of specific mechanisms of resistance (if

©Clinical and Laboratory Standards Institute. All rights reserved. 33

Copyright Property of CLSI for use with PK-PD FDA workshop.
Downloaded on 4/27/2017.
 M23, 4th ed.

known), and of virulence markers (if applicable to individual pathogens). In addition, MICs should be
displayed for organisms that are commonly grouped together for the purposes of determining
interpretive criteria (eg, Enterobacteriaceae).

900 875
Preclinical
800 Clinical
Number of Isolates

700
600 530
500
400 333
276
300
184 190
200 141

100 6 30 46
9 11 26 2 10 3 5
0
0.03 0.06 0.12 0.25 0.5 1 2 4 8
MIC (g/mL)

Abbreviation: MIC, minimal inhibitory concentration.

Figure 1. Distribution by MIC of Isolates Collected in Surveillance Studies and Clinical Trials

 The number of isolates tested for MIC determination in each laboratory should be reported if multiple
laboratories generated the datasets.

 A description of the MIC distributions by individual geographic region and all regions combined with
a discussion of any variability observed in light of prevalent resistance mechanisms.

 Minimum bactericidal concentrations should be submitted when appropriate based on mechanism of

action of the drug.

5.1.3.3 Cross-Resistance Between Antimicrobial Agents

Sponsors should compare the activity of a new antibacterial drug product to the activity profile of other
antimicrobial agents with the same or a very similar mechanism of action to assess the possibility of cross-
resistance. Sponsors should present and evaluate the potential for cross-resistance.

Under some circumstances, tentative inferences can be drawn about cross-resistance between antimicrobial
agents within a specific population of isolates from regression analyses of MICs for one antimicrobial agent
compared to another. If cross-resistance exists, a strong correlation between the MICs of the two agents
compared would be expected, and a majority of the MICs will be clustered on a 45-degree diagonal. If
resistance affects the activity of one agent more than the other, the regression line will fall above or below
the 45-degree diagonal (see Figure 2).

©Clinical and Laboratory Standards Institute. All rights reserved.

34
Copyright Property of CLSI for use with PK-PD FDA workshop.
Downloaded on 4/27/2017.
 M23, 4th ed.

Figure 2. Scattergram Comparing New Drug to Existing Drug. Vertical lines are presumptive
susceptible and resistant interpretive criteria for a new drug, and horizontal lines are susceptible and
resistant interpretive criteria for the existing drug. MIC correlates that are in the upper right and lower left
boxes have categorical agreement. The diagonal line marks equivalent MICs for the two drugs (absolute
agreement).

* MICs should be tabulated for the antimicrobial agent under study compared to one or more established
agents according to the phenotypic and/or genotypic characteristics of the isolates tested. For example, if
the antimicrobial agent under study is a new gyrase inhibitor, the MICs should be compared with those for
established fluoroquinolones against fluoroquinolone-resistant isolates; the methodology and the criteria
used to characterize isolates as resistant should be described.

* Determination of Epidemiological Cutoff Values

* ECVs should be determined for each organism or group of organisms for which interpretive criteria are
proposed.

ECVs are determined by collecting and merging MIC distribution data from a range of sources, and then
applying techniques for estimating the upper end of the wild-type distribution. In order to be reliable, ECVs
are estimated by accounting for both biological (strain-to-strain) variation and MIC assay variation within
and among laboratories. They are based on the assumption that the wild-type distribution of a particular
antimicrobial agent–organism combination does not vary geographically or over time. A number of
conditions must be fulfilled in order to generate reliable ECVs. The most important are:

 An ECV can only be determined within a single species because of the genetic diversity among species
within a genus.

 MIC values included in the merged dataset must have been determined using a recognized reference
method such as the CLSI MIC broth dilution method (refer to CLSI document M072), which is also the
international reference standard.21

 Data must be sourced from at least three separate laboratories, and there should be at least 100 unique
strains included in the merged dataset and weighted before pooling if more than 50% of MICs were
generated in a single laboratory.

 As much as possible, the MIC values included in an individual laboratory’s data must be on scale. This
applies particularly to MICs of the presumptive wild-type strains.
©Clinical and Laboratory Standards Institute. All rights reserved. 35
Copyright Property of CLSI for use with PK-PD FDA workshop.
Downloaded on 4/27/2017.
 M23, 4th ed.

 Before merging data for ECV estimation, the MIC distribution from each individual laboratory is
inspected, and if the lowest concentration tested is also a mode, then these data cannot be included
in the merged dataset.

Once acceptable data are merged, a number of methods can be used to estimate the ECV. The method used
to determine ECVs should be described.

The simplest method is that of visual inspection, which generally works for MIC distributions when there
is clear separation of wild-type and non-wild-type. When there is no clear separation between wild-type
and non-wild-type strains, visual inspection becomes very subjective and recourse to statistical methods
should be considered to remove any potential observer bias from the estimation. Several methods have been
proposed and published.22-27 The sponsor should explain and justify the methodology that is used.

Estimation of ECVs from MIC distributions may be supplemented with molecular tests for known
resistance mechanisms as a form a validation. The detection of a resistance gene per se in organisms with
MICs at or below the ECV does not necessarily invalidate the choice of ECV, unless it can be accompanied
by evidence that the gene is being expressed.

5.2 Nonclinical Pharmacokinetic-Pharmacodynamic Cutoff

The nonclinical PK-PD cutoff is the highest MIC value at which efficacy would be predicted in patients
based on nonclinically derived PK-PD relationships and human PK exposures. The nonclinical PK-PD
cutoff is derived from identifying the PK-PD target (which is defined by the PK-PD index best describing
the exposure-response relationship and the magnitude of the PK-PD index associated with efficacy) in
nonclinical studies, then applying it in Monte Carlo simulation using human PK to predict the percentage
of patients likely to achieve the PK-PD target at a given MIC.

Analyses describing the relationship between exposure and response for efficacy of antimicrobial agents in
nonclinical studies must be included in the submission. Standardized PK-PD terminology for anti-infective
agents must be used.28

 PK-PD index refers to the quantitative relationship between a measure of drug exposure, such as area
under the curve (AUC) and a microbiologic parameter such as MIC.

 PK-PD magnitude refers to the numerical value of the PK-PD index.

 PK-PD target is a magnitude for a PK-PD index at which a desired level of predicted response is
achieved.

The science of PK-PD relationships is an evolving one. Therefore, submissions to the CLSI Subcommittee
on AST should clearly describe the sources of data, assumptions, and details of the models and simulations
used to support the proposed interpretive criteria. This information will allow comparisons with other
methods and facilitate re-examination of nonclinical PK-PD cutoffs if it becomes necessary to do so in the
future.

Static In Vitro Studies

Static in vitro experiments, such as time-kill studies and postantibiotic effect (PAE) studies, serve as an
initial step in the process of identifying the PK-PD index or indices most closely associated with efficacy.
Static in vitro experiments can serve to support nonclinical and clinical PK-PD study results.

For antimicrobial agents that display a concentration-dependent pattern of bactericidal activity in time-kill
studies, the ratio of area under the concentration-time curve over 24 hours to the MIC of the pathogen
©Clinical and Laboratory Standards Institute. All rights reserved.
36
Copyright Property of CLSI for use with PK-PD FDA workshop.
Downloaded on 4/27/2017.
 M23, 4th ed.

(AUC0-24:MIC ratio) and/or the ratio of the maximal drug concentration to the MIC of the pathogen
(Cmax:MIC ratio) is usually predictive of efficacy in PK-PD model systems. For antimicrobial agents that
display a time-dependent pattern of bactericidal activity in time-kill studies without PAE, the percentage of
time during a dosing interval that drug concentrations remain above the MIC of the antimicrobial agent for
the pathogen (%Time > MIC) is usually predictive of efficacy in PK-PD model systems. For antimicrobial
agents that display a time-dependent pattern of bactericidal activity in time-kill studies and demonstrate
PAE, the AUC0-24:MIC ratio is usually predictive of efficacy in PK-PD model systems. For agents that
have intracellular activity, information on intracellular concentration and activity against target pathogens
may be appropriate.

Sponsors may provide:

 Time-kill study results, when such studies have been conducted

 At a minimum, studies should be conducted using isolates relevant to the intended clinical uses
(representative of the predominant pathogens and/or the major resistance mechanisms), and over a
range of clinically relevant drug concentrations. Using well-characterized strains, such as ATCC®
strains, is encouraged to allow comparison to historical data for other antimicrobial agents.

 Studies should be conducted at least in duplicate. The sponsor should follow standard procedures
described in CLSI document M26.29 If other methods are used, details of the procedures must be
provided.

 PAE study results, when such studies have been conducted

 PAE studies should be conducted with clinically relevant isolates.
 Studies should be conducted at least in duplicate. Detailed methods must be provided.30

Nonclinical Pharmacokinetic-Pharmacodynamic Studies

* Nonclinical PK-PD infection model systems are used to identify the PK-PD index or indices that are most
closely associated with efficacy. Additionally, nonclinical PK-PD model systems provide insight as to the
magnitude of the PK-PD index or indices necessary to achieve relevant clinical end points in patients. Data
from these studies are used to ultimately derive the nonclinical PK-PD cutoff. Nonclinical PK-PD studies
can serve to support clinical PK-PD study results when such data are available.

* The sponsor should conduct studies (or provide existing data) to identify the PK-PD index or indices most
closely related with efficacy. Indices typically evaluated include %Time > MIC, AUC0-24:MIC ratio, and
Cmax:MIC ratio. Other indices may also be evaluated, as appropriate. Typically, these data are generated
from adequately controlled and well-designed, nonclinical PK-PD infection models (animal or in vitro).
Examples of acceptable models include the neutropenic mouse thigh and lung infection models31,32 and in
vitro PD models33 (eg, chemostat or hollow-fiber systems). Studies should be designed with treated and
untreated control arms, and PK should be assessed. PK-PD data from these studies should be of high quality
with well-defined model conditions and study design variables. Detailed methods and pertinent
experimental data should be provided for appropriate interpretation of PK-PD results.

Other nonclinical models (eg, non-neutropenic mouse model) may be considered if sufficient data are
submitted to demonstrate correlation of the proposed model to an acceptable model or to the clinical
indication. Ideally, nonclinical models should mimic the human indication as closely as possible. This is of
particular interest for indications when drug penetration to the infection site is an issue (eg, endocarditis,
meningitis, and pneumonia).34 Whenever possible, an assessment of exposure to the antimicrobial agent at
the infection site (eg, epithelial lining fluid [ELF] for agents intended to treat pneumonia) in addition to

©Clinical and Laboratory Standards Institute. All rights reserved. 37

Copyright Property of CLSI for use with PK-PD FDA workshop.
Downloaded on 4/27/2017.
 M23, 4th ed.

plasma or serum is recommended. Site-specific infection models may be considered supplementary to the
acceptable PK-PD models mentioned above.

* When conducting nonclinical PK-PD studies, a range of suitable isolates of species representative of those
most relevant to the intended clinical uses and exhibiting an MIC range across the wild-type distribution
should be examined. A minimum of three strains should be tested for each bacterial species or group that
has its own interpretative criteria (see CLSI document M100S5). In the case of Enterobacteriaceae, testing
≥ 3 strains of each species most relevant to the clinical uses is appropriate. Strains with different phenotypes
(eg, methicillin-resistant S. aureus vs methicillin-susceptible S. aureus, extended-spectrum β-lactamase
[ESBL] vs non-ESBL) should be evaluated, if relevant. Including a few strains that have been previously
characterized in PK-PD models (eg, ATCC® strains) is also recommended to allow comparison to historical
data generated with other antimicrobial agents. Testing of additional isolates beyond the minimum
requirement may be necessary to assess for interstrain PD variability that is not explained by MIC alone.35

* When undertaking analyses of PK-PD relationships, model parameters and goodness of fit should be
stated. For example, when a Hill-type function is fit to PK-PD data, sponsors should provide E0, EC50,
Emax, and Hill’s constant. Reporting of the 95% confidence interval around the mean parameter values is
encouraged. Details of the analysis methods used should be provided.

* Each PK-PD index should be expressed as a function of free drug concentrations. If the PK-PD index or
indices are not presented as free drug, the sponsor must provide scientific reasoning to justify why this is
appropriate. The magnitude of the PK-PD index or indices necessary to achieve selected end points should
be provided. These may include EC80, net bacterial stasis, or 1-, 2-, and/or 3-log10 reductions in bacterial
densities. Although no specific end point is consistently most relevant, general guidance can be provided
using the neutropenic mouse thigh model, for which there is evidence demonstrating the translation to
humans.31,32 Some example scenarios include:

 For severe, life-threatening infections of high bacterial burden, such as nosocomial pneumonia,
achieving the magnitude of the PK-PD index associated with a minimum of a 1-log10 reduction using
the neutropenic mouse thigh or lung model has correlated with favorable clinical outcomes in patients.32

 For low-density infections that are treated in part by surgical interventions, such as complicated skin
and skin-structure infections or complicated intra-abdominal infections, achieving the magnitude of the
PK-PD index associated with a minimum of net stasis using the neutropenic mouse thigh model has
correlated well with favorable clinical outcomes in patients.32

* The sponsor should provide justification for the nonclinical PK-PD target selected to derive the
nonclinical PK-PD cutoff. The rationale should describe the magnitude of the PK-PD index that is being
targeted for the selected efficacy end point (ie, the PK-PD target) and how it relates to the clinical indication.

When interpretative criteria are being proposed for a species for which robust nonclinical models have not
been validated, it may be acceptable to extrapolate the nonclinical PK-PD target from one species to
another. The sponsor should provide evidence to support the relevance of the proposed nonclinical PK-PD
target to the species. Advice can also be requested from the CLSI Working Group on AST Breakpoints.

Human Pharmacokinetic Data

* Human PK data are critical for development of interpretive criteria for in vitro susceptibility testing. These
data:

©Clinical and Laboratory Standards Institute. All rights reserved.

38
Copyright Property of CLSI for use with PK-PD FDA workshop.
Downloaded on 4/27/2017.
 M23, 4th ed.

 Allow an understanding of exposure to the antimicrobial agent achieved following administration of

the selected dosing regimens in target patient populations (ie, the types of infected patients in whom
the antimicrobial agent is likely to be used).

 Permit the assessment of interpatient variability.

 Permit analysis of the covariates that influence drug exposure.

* Population PK models should be developed in order to predict human drug exposure for Monte Carlo
simulations and for analyses exploring exposure-response relationships in the target patient population.

* The sponsor should provide PK information derived from the target patient population. These data should
be sufficient (eg, derived from limited detailed and more extensive sparse sampling) to provide robust point
estimates of PK parameter values, assess interpatient variability, and identify clinically significant
covariates. If PK data from the target patient population are not available, the sponsor must provide
information from relevant volunteer and special population (eg, renal impairment) PK studies along with
an explanation as to why these data are adequate in place of target patient population data.

* The sponsor must provide information on human protein binding in order to estimate free drug
concentrations. Human protein binding should be evaluated at clinically relevant concentrations and over a
range to assess for concentration-dependent binding. Detailed methods must be provided.

If available, total and free drug concentration-time data in relevant nonhomogenate tissues and body fluids
should be presented to help understand drug penetration to the infection site (see examples below). Drug
penetration should be evaluated by the rate and extent of drug exposure at the body site relative to plasma
or serum. Although the science around PK-PD using body site PK is evolving, the difference in
the tissue:plasma or tissue:serum exposure ratio between humans and animals (from which the nonclinical
PK-PD target is derived) should at least be considered.36

Some examples of data showing drug penetration to the infection site include:

 Data showing the PK of the free drug in urine, which should be provided if the drug will be used to
treat urinary tract infection.

 Free drug concentration data in ELF, which may be important in the evaluation if antimicrobial agents
are intended to treat pneumonia.34 Studies to assess ELF penetration often use a population approach
with a single bronchoalveolar lavage assessment per subject for a compilation of multiple assessments
at different time points relative to the time of drug administration.

 Free drug concentration data in CSF, which must be presented if the antimicrobial agent is intended to
treat meningitis. CSF penetration studies often use a population approach for sampling.

 Free drug concentration data in vegetations, abscesses, and skin, which may be helpful for drugs that
are to be used to treat endocarditis, abdominal abscesses, and skin infections, respectively. For skin
infections, the data can be obtained with a skin blister fluid or tissue microdialysis study.

Monte Carlo Simulation and Probability of Target Attainment

5.2.4.1 Monte Carlo Simulation

* Monte Carlo simulation is useful for the evaluation of potential nonclinical PK-PD cutoffs.37 Typically,
simulation inputs include measures of central tendency statistics for PK parameters and their associated
variance. Through this statistical algorithm, it is possible to estimate the probability of attaining the target
©Clinical and Laboratory Standards Institute. All rights reserved. 39
Copyright Property of CLSI for use with PK-PD FDA workshop.
Downloaded on 4/27/2017.
 M23, 4th ed.

drug exposure consistent with efficacy (derived from nonclinical PK-PD studies) in the context of the MIC
of the pathogen(s) of interest, which is more formally known as the probability of target attainment (PTA).

A large variety of software programs can be used to conduct Monte Carlo simulation analysis. Monte Carlo
simulation has been the technique most commonly used to provide data for evaluation by the CLSI
Subcommittee on AST in determining nonclinical PK-PD cutoffs, and the information needed to support
these simulations is indicated below. If alternative methods are used by the sponsor, corresponding elements
should be assumed necessary.

* It is preferred that the PK inputs for Monte Carlo simulations be based on a target patient population PK
model. The validity of this tool is based upon an assumption that the population PK model provides an
adequate description of drug exposures in the target patient population. Although there is no set number of
individuals that should be included in developing a patient population PK model, generally the dataset
should be sufficient to provide precise estimates of the main PK parameters and their associated variances
and allow for assessment of the covariates’ effect.

* If a target patient population PK model is not available, the sponsor may use a relevant volunteer-derived
PK model. When using a population PK model constructed from healthy volunteer data, it is imperative to
make adjustments to the model so the Monte Carlo simulation results more accurately reflect drug exposure
and its associated variability in the target patient population. An explanation as to why the volunteer-derived
PK model is adequate in place of a target patient population model must be provided along with sensitivity
analyses.

Examples of adjustments to the population PK model include:

 In healthy volunteers, the mean point estimate of drug clearance is often somewhat similar, but the data
dispersion around the mean point estimate is considerably less variable than that observed in a patient
population. In such a circumstance, one should artificially inflate the variability around the point
estimate of drug clearance to a magnitude similar to that expected in the target patient population.
 The inflated variance approach is crucial for any drug that may be used in patients with severe
infections associated with considerable systemic upset because of the large PK variability observed
in this population.

 The level of variance applied for simulation should be justified by the sponsor.

 Healthy volunteers often have normal renal function whereas patients often have some degree of renal
impairment.
 In this scenario, one should impose a representative distribution for creatinine clearance to better
reflect that seen in the target patient population.

 The distribution of renal function applied for simulation should be justified by the sponsor.

* Monte Carlo simulation should be performed using the same PK model from which the PK parameter and
dispersion estimates were obtained. Exceptions are model adjustments, as previously described, intended
to better estimate PK and associated variability in the target patient population.

* Monte Carlo simulations can use PK-PD relationships derived from nonclinical infection models or from
the target patient population. When the nonclinical PK-PD target is used, differences in protein binding
between the nonclinical and human systems must be considered.

* Simulation results must be presented stratified by MIC value and in the context of the MIC distribution
for the species of interest. If any of the MIC values were censored (eg, values ≤ 0.25 µg/mL were changed

©Clinical and Laboratory Standards Institute. All rights reserved.

40
Copyright Property of CLSI for use with PK-PD FDA workshop.
Downloaded on 4/27/2017.
 M23, 4th ed.

to 0.125 µg/mL), this must be noted. For those Monte Carlo analyses performed using censored MIC
populations, simulations should be performed with and without those values.

* The underlying population distributions (eg, normal, log normal) and/or assumptions used for the various
inputs (PK data, MIC data) as well as the number of simulations performed, must be described. The total
number of iterates (ie, simulated patients) conducted in the Monte Carlo simulation is dependent on the
variability of the data and the complexity of the analyses. Generally, 1000 to 5000 iterates are sufficient for
the vast majority of Monte Carlo simulations.

5.2.4.2 Probability of Target Attainment

* PTA is the most commonly reported output for Monte Carlo simulation and is the final step in determining
the nonclinical PK-PD cutoff.

* For the purpose of identifying nonclinical PK-PD cutoffs, 90% PTA at a given MIC is considered
acceptable by the Subcommittee on AST. However, there are circumstances in which higher or lower
percentages may be acceptable. For example, if the consequence of not attaining the target PK-PD threshold
is a high probability of a patient’s death (eg, suboptimal drug exposure in the context of a patient with
pulmonary anthrax), higher PTA thresholds may be judged appropriate. Conversely, a lower PTA may be
acceptable if the infection is less severe.

Additional Considerations

* The sponsor must provide details of the methods used for measurements of total and active drug
concentration in animal and human plasma, serum, or body fluids. This material should provide a
description of any active metabolites of the antimicrobial agent to humans or animals.

The relative performance of assays for drug concentration measurement, such as high-performance liquid
chromatography, should be provided. If bioassay methods are used for antimicrobial agents with active
metabolites, relevant information about the differences in metabolite formation in humans and animals
should be provided.

5.3 Clinical Exposure-Response Cutoff

The CER cutoff is the highest MIC value at which efficacy would be predicted in patients based on CER
relationships for efficacy in infected patients and on human PK. In this document “exposure-response
relationship” is used to refer to the exposure-response relationship for efficacy; it does not include the
exposure-response relationship for safety.

The CER cutoff is derived by using a population PD model based on data from humans to describe drug
disposition, Monte Carlo simulation to generate an exposure distribution, and an exposure-response
relationship based on clinical data. Acceptable approaches for analyses that can be conducted to identify
exposure-response relationships based on clinical data and the application of such data to identify a CER
cutoff are described in Subchapter 5.3.1.

As discussed in Subchapter 5.3.1, deriving a CER cutoff may not be possible using data from the clinical
trial program, for a variety of reasons. Streamlined programs studying limited numbers of patients may
provide insufficient data to support the identification of such a relationship. Additionally, if high success
rates are observed in trials in conjunction with a dosing regimen that effectively provided all (or nearly all)
patients with an exposure that was high relative to the MIC of the antimicrobial agent for their pathogens,
this will reduce the likelihood of being able to identify a CER cutoff.

©Clinical and Laboratory Standards Institute. All rights reserved. 41

Copyright Property of CLSI for use with PK-PD FDA workshop.
Downloaded on 4/27/2017.
 M23, 4th ed.

That limitation understood, analyses describing the exposure-response relationships of antimicrobial agents
in infected patients should be included where feasible in submissions to the CLSI Subcommittee on AST
to support requests for interpretive criteria. Standardized PK-PD terminology for anti-infective agents must
be used when reporting such analyses, ie, PK-PD index, PK-PD magnitude, and PK-PD target (see
Subchapter 5.2).28,37 However, for determining a CER cutoff, where the exposure-response relationships
will be determined in infected patients vs animal models of infection, these parameters are referred to as
clinically derived.

NOTE: The science of determining exposure-response relationships in infected patients is evolving.

Therefore, submissions to the CLSI Subcommittee on AST should clearly describe the sources of data,
assumptions, and details of the models and simulations used to support the identification of the CER cutoff.
This information allows comparisons to be made with other methods and facilitates re-evaluation of CER
cutoffs should it become necessary to do so in the future.

Analyses to Identify Exposure-Response Relationships for Efficacy

The CER relationships are established based on data from an evaluable population of infected patients
enrolled in clinical trials in which:

 Clinical outcomes are assessed.

 Adequate PK data are collected.
 Baseline MIC values of the antimicrobial agent for the pathogens of interest are determined.

Sponsors need to provide justification for efficacy end points chosen for evaluation (see Subchapter 5.3).

Statistical approaches for evaluating univariable exposure-response relationships are based on the nature of
the variables for the efficacy end point and the PK-PD index to be evaluated. Acceptable approaches that
can be undertaken include those described in Table 5. Sponsors should provide justification for alternative
methods.

Table 5. Evaluating Univariable Exposure-Response Relationships

Efficacy End Point Category
PK-PD Index Dichotomous
Category Variables Continuous Variables Time-to-Event
Linear regression, Cox proportional
Logistic regression and hazards and/or
Continuous variables Spearman correlation,
Hill-type/Emax parametric regression
and Hill-type/Emax
models
Analysis of variance
Chi-square or Fisher
Categorical variables and Wilcoxon rank sum Log-rank tests
exact test
test
Abbreviation: PK-PD, pharmacokinetic-pharmacodynamic.

As shown in Table 5, PK-PD indices can be evaluated as continuous or categorical variables. The
assessment of PK-PD indices as categorical variables allows for potential nonlinearity or nonmonotonicity
to be characterized. To this end, a tool such as classification and regression tree (CART) analysis can be
used to provide guidance on how to construct two-group or multigroup categorical variables. For a
dichotomous efficacy end point, the resulting splits of a classification tree can be used; for a continuous
efficacy end point, the resulting splits of a regression tree can be used.

If other patient factors in addition to the PK-PD index are found to be predictive of the efficacy end point
based on the results of univariable analyses, multivariable analyses should be undertaken to evaluate
predictors of outcome. In such cases, it may be more appropriate to consider distributions for such patient
©Clinical and Laboratory Standards Institute. All rights reserved.
42
Copyright Property of CLSI for use with PK-PD FDA workshop.
Downloaded on 4/27/2017.
 M23, 4th ed.

factors in addition to those for PK parameters when conducting Monte Carlo simulations to assess model-
predicted percent probabilities of response.

Application of Exposure-Response Relationships to Identify the Clinical Exposure-Response

Cutoff

The exposure-response relationship identified, as described in Subchapter 5.3.1, can then be used to identify
the CER cutoff. Two typical methods are described in the following subchapters.37

5.3.2.1 Determination of a Clinical Exposure-Response Cutoff Based on the Model-Predicted Probability

of Response by Minimal Inhibitory Concentration

Using a population PK model based on data from humans to describe drug disposition, Monte Carlo
simulation to generate an exposure distribution, and the CER relationship, model-predicted percent
probabilities of response can be determined for a dichotomous efficacy end point at given MIC values. For
a continuous end point, mean (or median) response can be evaluated by MIC. In the case of a dichotomous
end point, the MIC chosen as the CER cutoff is determined by the selected summary statistic (eg, median)
which provides a percent probability of model-predicted response that is consistent with that which is
desired for the disease state.

5.3.2.2 Determination of a Clinical Exposure-Response Cutoff Based on the Probability of

Pharmacokinetic-Pharmacodynamic Target Attainment

Using a population PK model based on data from humans to describe drug disposition, Monte Carlo
simulation to generate an exposure distribution, and a clinically derived PK-PD target (eg, a CART-derived
value that is highly predictive of response, or a model-derived value associated with a high probability or
level of response), the percent probability of PK-PD target attainment can be determined at given (or
various) MIC values. For dichotomous efficacy end points, the MIC chosen as the CER cutoff is determined
as the highest MIC at which target attainment of at least 90% is achieved.

An important limitation of this approach is that failure to achieve a PK-PD target as evidenced by low
percent probabilities of PK-PD target attainment at a given MIC value does not necessarily predict that the
percentage of successful responses will be equally low at that same MIC value. Thus, the evaluation of
model-predicted percent probabilities of response described above, when possible to carry out, is expected
to provide predictions that are more aligned with observed response by MIC value than assessments of
percent probabilities of PK-PD target attainment.

Additional Considerations

Sponsors are strongly encouraged to collect adequate PK data in Phase 2 and 3 clinical trials for the purpose
of attempting to establish CER relationship(s) for efficacy so that a CER cutoff can be determined. The
exposure-response analyses performed should consider the nature of the clinical end point and whether
univariable or multivariable analyses are more appropriate. The sponsor should fully justify why a particular
analytical approach was used to identify a CER cutoff. Bayesian analysis of preclinical and clinical data
may be considered as an appropriate option.38 Exposure-response relationships derived from clinical data
may not be sufficiently optimal to set an initial CER cutoff. This is commonly due to inability to adequately
characterize exposure-response relationships for efficacy based on data from Phase 2 and 3 clinical trials in
which too few failures occurred or there were too few patients with exposures in a critical range. However,
some recommendations to enhance the likelihood of generating sufficient data to support a CER cutoff are
made in the following paragraphs.

The number of patients needed for a meaningful exposure-response analysis is influenced by the dose range
under study, the sample size, and characteristics and quality of the data. For example, for a dichotomous

©Clinical and Laboratory Standards Institute. All rights reserved. 43

Copyright Property of CLSI for use with PK-PD FDA workshop.
Downloaded on 4/27/2017.
 M23, 4th ed.

efficacy end point, if the numbers of failures or successes are too high, it would be difficult to identify an
exposure-response relationship. For a time-to-event efficacy end point, the statistical approaches are more
sensitive so that more can be done with data from fewer patients as long as the percentages of censored data
are within reasonable limits. Specific guidance on numbers of patients needs to be provided in relation to
the actual study design. Time-to-event efficacy end points (eg, time to fever reduction) that have been
demonstrated to be clinically relevant as well as relevant to the drug exposure should be considered in study
designs.

Sponsors should report the diagnostics of the fitting of exposure-response data to statistical models (model
building) and the evaluation of the predictability of the model (model validation). These diagnostics lend
confidence to the strength of the underlying exposure-response relationship given the important role of this
relationship in the Monte Carlo simulations.

Although a CER cutoff is an attractive concept due to its integration of data from several sources and
obvious clinical relevance, it is not necessarily more reliable than the methods determining a cutoff based
on nonclinically derived PK-PD targets (eg, from animal models) and human PK exposures. Nevertheless,
if clinical data are sufficient to support a significant CER relationship, then the CER cutoff can be a useful
tool for predicting efficacy.

5.4 Clinical Cutoff

For the purpose of setting interpretive criteria, the clinical cutoff is based on a simple evaluation of any
correlation between clinical outcome and the MIC(s) of an antimicrobial agent(s) for the infecting
pathogen(s). See Subchapter 5.3 for a description of the CER cutoff, which is defined as the highest MIC
value at which likely efficacy would be predicted in patients based on CER relationships for efficacy in
infected patients and on human PK.

During the clinical evaluation of new antimicrobial agents, or in trials that evaluate an existing antimicrobial
agent for the treatment of new types of infections, the trials’ sponsor should investigate any correlation that
might be discernible between in vitro susceptibility test results and therapeutic outcomes (clinical and
microbiological). For established antimicrobial agents, sponsors proposing new or revised interpretive
criteria may not have access to the raw data of clinical trials. Thus, sponsors and the CLSI Subcommittee
on AST may have to rely on the limited information available from the literature and published regulatory
reviews.

A clinical cutoff based on a simple correlation of outcome to MIC has significant limitations, which the
CER cutoff attempts to overcome. Although it is still considered useful to review a simple outcome to MIC
data, it is not usually possible to determine clinical cutoffs from clinical trial data due to several factors,
which may include:

 The inherent structure of clinical trials, with exclusion of patients with resistant organisms.

 The rarity of pathogens resistant to the antimicrobial agent when preregistration trials are conducted.

 The limited range of infection types studied, which limits the microbial species treated.

 The contribution of other intrinsic and extrinsic factors (eg, host immune systems, adjunctive
treatments, adjunctive surgery) that contribute to between-patient variability.

 For some infections, the identification of the major pathogen may not be straightforward (eg, in
polymicrobial intra-abdominal infections), which limits the interpretation of analyses of outcomes by
patient or pathogen.

©Clinical and Laboratory Standards Institute. All rights reserved.

44
Copyright Property of CLSI for use with PK-PD FDA workshop.
Downloaded on 4/27/2017.
 M23, 4th ed.

 Usually, the dose regimen(s) studied will have been chosen based on PK-PD analyses to ensure
coverage of organisms with MICs across the entire wild-type range. In the absence of significant
numbers of resistant organisms with very high MICs, the clinical outcome data (simple correlation of
outcomes to MIC) may not be able to provide meaningful evidence to either support or reject the
nonclinical PK-PD cutoffs and/or the CER cutoffs.

* Nevertheless, subject to availability of sufficient relevant data to the sponsor, submissions to the CLSI
Subcommittee on AST to support interpretive criteria should include a summary of the clinical data that
describes clinical and microbiological outcomes according to the MICs of the antimicrobial agent under
study that were documented for the infecting pathogens. The presentations of outcomes by MIC should be
separated by the dose regimen administered when clinical dose-finding studies were conducted.

In many cases, especially for new antimicrobial agents, the same data will have been submitted or will be
submitted to regulatory agencies. The sponsor should note that if the data presented to the CLSI
Subcommittee on AST differ from those presented to the regulatory agencies, the differences should be
described and explained.

The presentation of the clinical outcome data in the submission should include the elements described in
Subchapters 5.4.1 and 5.4.2.

Description of Clinical Trials

* The sponsor should provide a summary of the clinical trials. Each trial should be summarized with
attention to the following elements:

 Description of the study population

 Types of infection treated with details of diagnostic criteria

 Dose regimens for the antimicrobial agent under study and all comparative regimens, including any
switches that were allowed, with details of issues such as infusion times and total duration of therapy

 Timing of all visits at which clinical and microbiological data were collected, noting at which visits
outcomes were assessed

 Definitions of the analysis populations (eg, modified intent to treat, clinically evaluable,
microbiologically evaluable)

 Definitions of response

 Primary and secondary end points

 Statistical approach and sample size determination

Presentation of Outcomes

* Whenever the sponsor has access to the data, the presentation should include:

 Correlations between clinical and microbiological outcomes (categorized as in the protocol), drug
exposures, and actual MICs of the pathogens, analyzed by patient and by pathogen for individual
clinical trials and after pooling across trials

©Clinical and Laboratory Standards Institute. All rights reserved. 45

Copyright Property of CLSI for use with PK-PD FDA workshop.
Downloaded on 4/27/2017.
 M23, 4th ed.

– For the outcomes by pathogen, an analysis by genus (Escherichia, Klebsiella, Proteus) and by
individual species, for the more commonly encountered pathogens, should be included in addition
to analyses by broad category (eg, Enterobacteriaceae).

 Results, which should be obtained using established CLSI reference methods for the antimicrobial agent
– If any results are obtained using other methods or systems, then data must be presented to
demonstrate the comparability of such methods to CLSI reference methods (see Chapter 2).

 QC data to support the reliability of the MIC data

 The data described in the first bullet should be repeated for each of the predefined patient populations
for the analysis of efficacy

 The summarized approach to data handling, with special regard to the handling of missing data and
default of indeterminate outcomes to failures in some analysis populations

 Separate analyses, if applicable, for patients with infections at specific body sites and for patients with
concurrent bacteremia

 A detailed consideration of the data from patients with clinical and/or microbiological failure

 A separate presentation of any data on outcomes from patients who were treated despite having an
infection due to an organism that would be designated resistant based on the proposed interpretive
criteria

5.5 Consideration of the Cutoffs to Determine Interpretive Criteria

Subchapters 5.1, 5.2, 5.3, and 5.4 describe the:

 ECV
 Nonclinical PK-PD cutoff
 CER cutoff
 Clinical cutoff

For the purposes of determining susceptibility test interpretive criteria, it is necessary to take into account
and to weight the various sets of data and outputs of the relevant analyses that may be available. It is unlikely
that a submission to the CLSI Subcommittee on AST to support determination of interpretive criteria will
include all the above elements and cutoffs. For example, available data may be inadequate to support
derivation of a CER cutoff. In addition, it is very unusual that analyses of clinical outcomes by MIC can
yield a clear clinical cutoff.

Depending on the available data and the cutoffs that have been identified, the combinations of information
relevant to determining interpretive criteria will vary considerably. Some approaches to consideration of
the available data include the following scenarios in which it is assumed that any differences there may be
between available cutoffs are ≤ 2-fold. If the difference between any available cutoffs is > 4-fold, the
possible reasons for the discrepancy must be explored before any attempt is made to determine interpretive
criteria from the cutoff values.

 If the ECV for a particular species or group of species is significantly greater than the nonclinical PK-
PD and the CER and/or clinical cutoffs, then it can be safely assumed that the species is intrinsically
resistant to the antimicrobial agent when used at the dosing regimen on which the cutoffs are based.

©Clinical and Laboratory Standards Institute. All rights reserved.

46
Copyright Property of CLSI for use with PK-PD FDA workshop.
Downloaded on 4/27/2017.
 M23, 4th ed.

 If the ECV is less than the other cutoffs, then it is clear that the species is intrinsically susceptible, and
it is only necessary to choose among the available cutoffs, if they are not the same, in order to select
interpretive criteria.

 If the available cutoffs are not the same, the relative weighting of the cutoffs will depend on their
perceived robustness. For example, if a clinical cutoff was established, this is likely to be less reliable
than the nonclinical PK-PD cutoff. If a CER cutoff was established using a full pharmacometric analysis
of clinical trials, but it differs from the nonclinical PK-PD cutoff, it may be more difficult to give one
more weight over the other. The value of the data supporting each cutoff would need to be evaluated
on a case-by-case basis.

 There is more uncertainty when one or more of the available cutoffs fall within the wild-type
distribution of MICs. In general, it is considered undesirable to set interpretive criteria that fall within
the wild-type MIC distribution because the interpretation of the test result will be greatly influenced by
normal test variation. When this situation occurs, there is a need to consider the potential clinical effect.
For example, if the intended use of the antimicrobial agent is for nonsevere infections caused by the
species or group of organisms being examined, and the preferred interpretive criteria based on available
cutoffs fall near the high end of the wild-type distribution, then there may not be a significant effect on
appropriate clinical use.

 If there are only an ECV and a nonclinical PK-PD cutoff, then the considerations are similar to those
outlined above in the third and fourth bullet points.

 It is expected to be very unusual that a submission will include only an ECV and a clinical cutoff. On
occasion, this situation may occur if a revision of interpretive criteria is proposed for an antimicrobial
agent that has been in use for many years and signals from clinical studies or during routine use suggest
that the current interpretive criteria are inappropriate, and if there is perceived to be sufficient urgency
to revise the criteria without generating data to support a nonclinical PK-PD cutoff.

©Clinical and Laboratory Standards Institute. All rights reserved. 47

Copyright Property of CLSI for use with PK-PD FDA workshop.
Downloaded on 4/27/2017.