2010 dose-effect modeling of experimental data full

Upload: -

Post on 06-Apr-2018

218 views

Category:

Documents


0 download

TRANSCRIPT

  • 8/3/2019 2010 Dose-Effect Modeling of Experimental Data Full

    1/11

  • 8/3/2019 2010 Dose-Effect Modeling of Experimental Data Full

    2/11

  • 8/3/2019 2010 Dose-Effect Modeling of Experimental Data Full

    3/11

  • 8/3/2019 2010 Dose-Effect Modeling of Experimental Data Full

    4/11

    Journal of Information, Control and Management Systems, Vol. 8, (2010), No.3 257

    DOSE-EFFECT MODELING OF EXPERIMENTAL DATA

    Kaloyan YANKOV

    Medical Faculty, Trakia University,

    Stara Zagora, Bulgaria

    [email protected]

    Abstract

    This paper proposes a mathematical model for estimation of the dose-response

    relationship of experimental data. modified logistic function is used as

    analytical model for determining the dose-effect. Parameters of the model are

    identified with the optimization procedure based on the cyclic coordinatedescent method. Formulas are derived to calculate the effective dose level and

    the standard deviation of doses. The approach is implemented in the computer

    program KORELIA-Dynamics.

    Keywords: dose-time-response, toxicity, effective dose level, lethal dose, modeling,coordinate descent method.

    1 INTRODUCTIONTo analyze the effects of a drug, the dose-response relationship is studied in

    pharmacology and toxicology. Dose-response experiments are routinely conducted in

    preclinical and clinical trials to study the relationship between the dose level of a drug

    and the probability of a response, be it cured or poisoned. Therefore the response

    of these experiments is binary or quantal, which means either the drug takes effect and

    the subject has a reaction or not. The underlying assumption for quantal bioassay is

    that a subject is administered a stimulus at a dose level x, and that the response Yis abinary random variable with success probabilityp(x), i.e.

    Y ~ Bin(1,p(x)).

    The function ]1,0[: p is called a dose response curve and is usuallyassumed to be strictly monotone. This curve describes the probability of success of the

    analyzed drug.

    For a given ))1(),0(( pp the parameter of interest is the effective dose level:

    ED = p1()

  • 8/3/2019 2010 Dose-Effect Modeling of Experimental Data Full

    5/11

    Dose-Effect Modeling of Experimental Data258

    This is the dose, where 100 % of the subjects reacts. In case the main effect is

    death, a lethal dose LD is defined.

    Depending on the value of

    several doses of practical importance are defined. = 0. Effective/lethal dose ED0/LD0 - the estimated dose at which none of

    the population is expected to react or die.

    = 1. One hundred percent effective/lethal dose ED100/LD100 the lowestpossible dose that causes effects in all studied cases.

    = 0.5. Median effective/lethal dose ED50/LD50 - the dose of a drug predicted to produce a characteristic effect in 50 percent of the subjects towhom it is given. The ED50 is the most frequently used standardized dose by

    means of which the potencies of drugs are compared. An ED50 can be

    determined only from data involving all or none (quantal) response.Further in the text the term used is effective dose, being understood that the

    results will apply also to the lethal dose.

    Experimental animals are treated with the drugs in order to study the effects and

    to determine the value of the base dose. The experiment is conducted under n dose

    levels. For each dose levelxi (i = 1,2,n),Nianimals are examined. They are observed

    over a period of time Tsufficient for the investigated drug to demonstrate its action

    and at the end of the experiment the number of reactived animals zi is noted. The

    probability that an individual animal on dose levelxiresponds by time Tis defined as:

    pi = zi/ NiIt is obvious that 0 = p0 p1 pn pn+1 = 1.

    The function presented by arranged pairs of points (xi , pi) is cumulative

    distribution function. It belongs to the class of the so-called S-shaped curves [1].

    Many publications deal with the problem of calculating the effective dose. The

    approaches for calculation of dose-effect based on the cumulative curve used in most

    of them are described below.

    1.1 Linear regression models.

    Linear interpolation. The effects determined during the experiment of theapplied doses are associated with line segments. The desired effective dose is

    calculated taking into account the coordinates in the relevant line segment.

    In case of explicit non-linearity a logarithmic scaling is performed upon thedata. Thus the graph acquires a relatively linear form, and then linear

    interpolation is applied. This approach uses the assumption that there is a

    logarithmic correlation between doses and effect of doses. For the entire range

    of applied doses that is not true, but in a narrow range around the relevant dose

    the method is appropriate.

    These two approaches were used before the era of modern computers when the

    most powerful computational tools were a pocket calculator and graph paper for

  • 8/3/2019 2010 Dose-Effect Modeling of Experimental Data Full

    6/11

    Journal of Information, Control and Management Systems, Vol. 8, (2010), No.3 259

    drawing graphs. Publications dealing with these methods are available from the 1970s

    till now [2,3,4].

    1.2 Nonlinear regression models. Interpolation of the data with cubic spline [5,6,7]. Piece-wise spline

    inperpolation is a convenient tool for interpolation of data being continuously

    differentiable to the second derivative. When appropriate software is

    available, this approach is very convenient and has great precision..

    Modeling techniques based on the analytical form of mathematical function.They assume a functional relationship between dose and response,

    according to a pre-specified parametric model, such as a linear, quadratic,

    logarithmic, exponential, logistic and so on [8]. By the application of

    minimizing numerical procedure, parameters of the function are determined,i.e. data identification is implemented. This approach provides the greatest

    accuracy due to the use of computers and specialized software. Another

    benefit of this approach is the additional parameters that the software can

    produce and the possibility for graphical illustration of the results. The most

    commonly used optimization methods are the method of least squares, the

    method of weighting by least squares or the maximum likelihood method [9].

    This paper aims to formulate a mathematical model of dose-response relationship

    using experimental data. The goal is to model the curve in the domain of definition and

    to minimize the possible error of approximation. The model is derived applying

    algorithm and software for system identification developed by the author [10,11,12].

    2 MODEL STRUCTUREOne of the possible equations for describing a sigmoidal curve is the Verhulst

    Pearl equation:

    )()(

    1)(

    xpK

    xpr

    dx

    xdp

    (1)

    p(x0) = C0Where: r the rate constant of the process

    Kthe asymptotic limit of growth (carrying capacity).

    This equation express quantitative changes in the presence of limited resources,

    represented by the constant .

    The solutionp(x) of Eq.(1) is the logistic function:

    http://www.ncbi.nlm.nih.gov/sites/entrez?Db=pubmed&Cmd=Search&Term=%22Huang%20Y%22%5BAuthor%5D&itool=EntrezSystem2.PEntrez.Pubmed.Pubmed_ResultsPanel.Pubmed_RVAbstractPlushttp://www.ncbi.nlm.nih.gov/sites/entrez?Db=pubmed&Cmd=Search&Term=%22Huang%20Y%22%5BAuthor%5D&itool=EntrezSystem2.PEntrez.Pubmed.Pubmed_ResultsPanel.Pubmed_RVAbstractPlus
  • 8/3/2019 2010 Dose-Effect Modeling of Experimental Data Full

    7/11

    Dose-Effect Modeling of Experimental Data260

    ).(

    011

    )(xr

    eC

    K

    Kxp

    (2)

    C0 = p(0)

    This equation may be revised in view of the goal to identify an effective dose,

    thereby makingK=1. o avoid the case C0=0 in Eq.(2), s training logistic function as

    proposed in [11] is used:

    )*(1

    1)(

    xrAexp (3)

    A>0, r > 0

    Three parameters are necessary to specify the curve, A, r and , defined as

    follows:

    A is the number of times that the initial valuep(0) must grow to reach 100%.

    ris the slope or the growth rate parameter that specifies "width" or "steepness" of

    the S-curve.

    is a dose correction parameter.

    These parameters must be identified by fitting the Eq.(3) to the experimental data.

    3 DATA IDENTIFICATIONTo fit the functionp(x) to the data, an identification procedure will be performed.

    The values pi in n number of points are known as a result of the measurement. They

    are used to identify the parameters,r, .

    Let us chose to subject to system identification the function

    ixrip

    AerAxp

    i

    )*(1

    1),,,(

    (4)

    A > 0, r > 0, i=1,2,,n

    The identification vectorQ(A,r,) must be defined from the system (4). But this

    system is not well defined. For each pointxi the residual di is:

    di(Q) = || pi-p(xi,Q) ||

    The aim is to minimize di(Q) for the whole interval of identification [xn-x1]:

    D(Q) = inf|| di(Q) || < , >0 (5)

    Thus the identification goal is translated into an optimization problem. The

    optimization method is defined depending on the way of minimization ofD(Q).

    Least square fitting.

  • 8/3/2019 2010 Dose-Effect Modeling of Experimental Data Full

    8/11

    Journal of Information, Control and Management Systems, Vol. 8, (2010), No.3 261

    An object of minimization is the functional:

    (6)

    uniform fitting is aimed at minimizing the maximal deviation in the base

    pointspi:

    min)(1

    2

    n

    i i

    dQD

    Uniform fittingThe

    min||max)( idQD (7)

    The Cyclic coordinate descent (CCD) method is an iterative heuristic search

    technique that attempts to minimize the error by varying one parameter at a time [10].

    It reduces the multi-criteria optimization to the single-criteria one. A number of passes

    are made over the model equation to find the global minimumD for system (5).

    4ssary to find the inverse function

    p-1

    (x). After trivial transformations of Eq.(3):

    CALCULATION OF DOSE-EFFECT PARAMETERS

    To calculate the effective or lethal dose, it is nece

    )(1 pr

    As a result of the identif

    )(

    lnln1

    )(p

    Ax (9)

    ication, the vector Q(A,r,) is known and x() is

    calcuFor most commonly used dose ED50 (LD50) atp()=0.5:lated directly from Eq.(9).

    Ar

    The final result is easy for computing. The

    xED ln1

    )5.0(50 (10)

    main burden is the process of

    ident

    uencyf(x) of

    the effectiveness of doses analytically and graphically (fig.1) is obtained:

    ification of these three parameters.

    The received analytical model of the sigmoidal curve allows obtaining other

    characteristics of dose-effect curve. After differentiation of Eq.(3) the freq

    2)(

    )(

    ]1[ Aedx

    The extreme point correspo

    )()(

    rx

    rxArexdpxf (11)

    nds to ED50/LD50. Thus the dose value of x(0.5) can

    be ca 0)( dxxdf .lculated from the equation

    Solution of the equation

    0)(

    2

    dx

    2

    xfd

  • 8/3/2019 2010 Dose-Effect Modeling of Experimental Data Full

    9/11

    Dose-Effect Modeling of Experimental Data262

    gives the coordinates of inflexion pointsIx1, Ix2 :

    rx1

    AI

    32

    ln

    rx2

    AI

    32

    ln

    and they are used to calculate the standard deviation of the frequency of doses (fig.1):

    rr

    IISD xx

    317.1

    2

    32

    32ln

    2

    12

    (12)

    It is possible to define three dose ranges: Low response dose. These doses are

    n range

    dose. Doses greater than

    (x(0.5)+2SD)

    Figure.1.Dose frequency

    5

    stand

    on linear interpolation to determine LD50

    . After identification with KORELIA-Dynamics the

    calcu

    lower than (x(0.5)-2SD)

    Mean response dose. Doses i(x(0.5)-2SD,x(0.5)+2SD)

    High response

    EXAMPLE

    The method is implemented in program KORELIADynamics using its ability for

    system identification. A dialog window is appended for calculation of ED50/LD50,

    ard deviation and visualization of sigmoidal curve and frequency of doses.

    The example from page 44 of [2] will be considered. The author discusses a

    toxicity of antibiotic tubazid. The calculations are performed using the Behrens

    approach (cit. Behrens B., Arch.exper. Path. Pharm., 140, 237, 1929) and oerber (cit.

    oerber G., Arch.exper. Path. Pharm., 162, 480, 1931). The experimental data are

    scattered on fig.2. Both methods are based

    and standard deviation SD. Cited data are:

    Behrens method: LD50SD = 172.3 9.5 [mg/kg]. oerber method: LD50SD =172.5 15 [mg/kg].

    The scatterplot on fig.2 should clearly indicate the appropriateness of using a

    logistic model (3) to fit this data

    lated model parameters are:

    A = 927.3; r = 0.1855; = -25.13

  • 8/3/2019 2010 Dose-Effect Modeling of Experimental Data Full

    10/11

    Journal of Information, Control and Management Systems, Vol. 8, (2010), No.3 263

    Identification errors are: absolute = 0.0346; relative = 0.0834; quadratic =

    0.00453.

    The identified equation for data is:

    )13.25*1855.0(

    1)(

    3.9271

    xxp

    e

    LD50SD = 172.3034 7.0997 [mg/kg].

    On fig.2 the experimental data, identification curve and frequency of doses are

    presented.

    Thus, the LD50 and SD according to equations (10) and (12) are

    Figure.2. Graph of the identification curve and frequency of doses

    ure that targets minimal residual using least-squa ulas for effective dose and

    ts the

    t the range of tested doses;

    Drops some requirements as strictly regulated modification of doses toconduct the experiment at pre-selected law;

    6 CONCLUSIONSThis work proposes an analytical model for determining the dose-effect.

    modified logistic function is used for data identification. Parameters of the model arecalculated with the optimization proced

    re or uniform criteria. From the analytical model form

    standard deviation of doses are derived.

    The advantages of the proposed analytical model are:

    Explicit provides the frequency of doses and standard deviation. Can provide the smallest error of approximation, because it reflec

    nonlinearity of the process by selecting the appropriate function;

    Shows the behavior of the system throughou

  • 8/3/2019 2010 Dose-Effect Modeling of Experimental Data Full

    11/11

    Dose-Effect Modeling of Experimental Data264

    A mathematical model can easily be embedded in more complex models,including the use of computers and also to broaden the community of research

    and modeled processes.

    The computer program Korelia-Dynamics offers friendly user interface for data

    input and graphical output of results. The program is easy and rapid to use, requiring

    minimum computer knowledge.

    REFERENCES:

    [1] KUCHARAVY, D., R. DE GUIO. Application of S-Shaped Curves, in 7th ETRIATRIZ Future Conference. Kassel University Press GmbH, Kassel: Frankfurt,

    Germany 2007.

    [2] BELENKII, M.A., Elements of Quantitative Evaluation of PharmacologicalEffects. GIML, Leningrad 1963 (russian).

    [3] SEPETLIEV, D., Statistical methods for data processing from medical researchstudies. dicina I fizkultura, Sofia, 1965 (bulgarian).

    [4] BRENNER D.J, L.R. HLATKY, P.J. HAHNFELDT, Y.HUANG, R.K.SACHS.The Linear-Quadratic Model and Most Other Common Radiobiological Models

    Result in Similar Predictions of Time-Dose Relationships. RADIATION

    RESEARCH 150, 83-91, 1998.

    [5] REMMENGA M. D., G. A. Milliken, D. Kratzer, J. R. Schwenke, H. R. Rolka.Estimating the maximum effective dose in a quantitative dose-response

    experiment. Journal of Animal Science, Vol 75: 2174-2183, 1997.

    [6] YANKOV, K., Software Utilities for Investigation of Regulating Systems, NinthNat. Conf. "Modern Tendencies in the Development of Fundamental and Applied

    Sciences". June, 5-6, 1998, Stara Zagora, Bulgaria, 401-408.

    [7] YANKOV, K. Evaluation of Some Dynamics Characteristics of TransientProcesses. Proc. 12-th Int.Conf. SAER'98. St.Konstantin resort, sept.19-20, 1998,

    Varna, Bulgaria. 113-117.

    [8] BRETZ F., Jason Hsu , Jose Pinheiro, Yi Liu. Dose Finding A Challenge inStatistics. Biometrical Journal 50 (4), 480504, 2008.

    [9] HUANG Y. Robustness of choice of number of doses for maximum likelihoodestimation of the ED(50) in bioassay. Stat Med. Aug 15;21(15):2215-23, 2002.[10] YANKOV, K. System Identification f Biological Processes. Proc. 20-thInt.Conf. "Systems for Automation of Engineering and Research (SAER-2006).

    St.St. Constantine and Elena resort, sept.23-24, Varna, Bulgaria, 144-149, 2006.

    [11] YANKOV, K Recognition and Function Association of Experimental Data.Proc. Int. Conference on Information Technologies (InfoTech-2009). Constantine

    and Elena resort, sept.17-20, Varna, Bulgaria, 131-140, 2009.

    [12] YANKOV, K Decision Planning of System Identification. Proc. Int. Conferenceon Information Technologies (InfoTech-2010). Constantine and Elena resort,

    sept.16-18, Varna, Bulgaria, 229-238,2010.

    http://www.ncbi.nlm.nih.gov/sites/entrez?Db=pubmed&Cmd=Search&Term=%22Huang%20Y%22%5BAuthor%5D&itool=EntrezSystem2.PEntrez.Pubmed.Pubmed_ResultsPanel.Pubmed_RVAbstractPlushttp://www.ncbi.nlm.nih.gov/sites/entrez?Db=pubmed&Cmd=Search&Term=%22Huang%20Y%22%5BAuthor%5D&itool=EntrezSystem2.PEntrez.Pubmed.Pubmed_ResultsPanel.Pubmed_RVAbstractPlus