AMD - PKPD

Will develop the best PKPD model based on a PK model.

The principle behind the PKPD model developement follows PPP&D, which initially will fixate the PK part of the model and only develope the PD part.

A complete PKPD workflow is hence currently only possible by first filtering the model dataset for PK information and running AMD for a ‘basic_pk’ model type. Then, given the resulting model from that workflow, the original dataset can be attached and another AMD run with model type ‘PKPD’ can be run. This can be done using pharmpy.modeling.set_dataset().

Regarding the dataset, two DVIDs are required.

DVID	Description
1	Assumed to be connected to the PK part of the model
2	Assumed to be connected to the PD part of the model

Note

Please note that it is only possible to run the AMD tool for the PD part of PKPD models. The tool expects a fully build PK model as input. If a dataset is inputed, a basic PK model is created and used as input, which is not recommended

Running

The code to initiate the AMD tool for a PKPD model:

PythonR

from pharmpy.modeling import read_model
from pharmpy.tools read_model, run_amd

start_model = read_model('path/to/model')

res = run_amd(
            modeltype='pkpd',
            input=start_model,
            search_space='DIRECTEFFECT(*)',
            b_init=0.1,
            emax_init=1.0,
            ec50_init=0.1,
            met_init=0.4
)

Arguments

The AMD arguments used for PKPD models can be seen below. Some are mandatory for this type of model building while others are optional, and some AMD arguments are not used for this model type. If any of the mandatory arguments is missing, the program till raise an error.

Mandatory

Argument	Description
input	Start model object. See input in amd
modeltype	Set to ‘pkpd’ for this model type.
b_init	Initial estimate for the baseline effect
emax_init	Initial estimate for the Emax
ec50_init	Initial estimate for the EC50
met_init	Initial estimate for the mean equilibration time

Note

In addition to these arguments, common arguments can be added.

Strategy components

For a description about the different model building strategies in AMD, see Strategy. This section will cover the aspects that are specific to PKPD models.

Structural

Structsearch

For a PKPD model, structsearch is run to determine the best structural model. All input arguments are specified by the user when initializing AMD. For more information regarding how the search space is used in structsearch, please see structsearch tool

Argument	Setting
search_space	search_space (As defined in AMD options)
modeltype	‘pkpd’
b_init	'b_init' (As defined in AMD input)
emax_init	'emax_init' (As defined in AMD input)
ec50_init	'ec50_init' (As defined in AMD input)
met_init	met_init (As defined in AMD input)
strictness	strictness (As defined in AMD input)

If no search space is given for the structsearch tool, then a default will be set to:

DIRECTEFFECT([LINEAR, EMAX, SIGMOID])
EFFECTCOMP([LINEAR, EMAX, SIGMOID])
INDIRECTEFFECT([LINEAR, EMAX, SIGMOID], *)

IIVSearch

The settings that the AMD tool uses for this subtool can be seen in the table below.

Argument	Setting	Setting (rerun)
algorithm	‘top_down_exhaustive’	‘top_down_exhaustive’
iiv_strategy	‘pd_fullblock’	‘no_add’
rank_type	‘mbic’ (type: iiv)	‘mbic’ (type: iiv)
cutoff	None	None
keep	Clearance parameters from input model	Clearance parameters from input model

Residual

The settings that the AMD tool uses for this subtool can be seen in the table below. When re-running the tool, the settings remain the same.

Argument	Setting
dv	2
groups	4
p_value	0.05
skip	None

IOVSearch

The settings that the AMD tool uses for this subtool can be seen in the table below.

Argument		Setting
column		occasion (As defined in AMD options)
list_of_parameters		None
rank_type		‘bic’ (type: random)
cutoff		None
distribution | ‘same-as-iiv’

Allometry

Allometry is completely skipped when running AMD with a model type of ‘pkpd’

covsearch

The settings that the AMD tool uses for this subtool can be seen in the table below.

Argument	Setting
search_space	search_space (As defined in AMD options)
p_forward	0.05
p_backward	0.01
max_steps	-1
algorithm	‘scm-forward-then-backward’

If no search space for this tool is given, the following default will be used:

COVARIATE?(@PD_IIV, @CONTINUOUS, exp, *)
COVARIATE?(@PD_IIV, @CATEGORICAL, cat, *)

Here, both statements are defined with a ‘?’, meaning that these are covariate effect(s) to be explored rather than structural covariate effects.

Structural covariate effects found in the search space are also added to the model in this step.

Mechanisitic covariates

If any mechanistic covariates have been given as input to the AMD tool, the specified covariate effects for these covariates is run in a separate initial covsearch run when adding covariates. These covariate effects are extracted from the given search space

Exploratory covariates

The remaining covariate effects from the search space are now run in an exploratory search.

Examples

Minimal

A minimal example for running AMD with model type PKPD:

PythonR

from pharmpy.tools import run_amd

start_model = read_model('path/to/model')

res = run_amd(
            modeltype='pkpd',
            input=start_model,
            b_init=2.0,
            emax_init=1.0,
            ec50=0.1,
            met_init=2.1
)

Specifying search space

PythonR

from pharmpy.tools import run_amd

start_model = read_model('path/to/model')

res = run_amd(
            modeltype='pkpd',
            input=start_model,
            search_space = "DIRECTEFFECT(linear)",
            b_init=2.0,
            emax_init=1.0,
            ec50=0.1,
            met_init=2.1
)